A multi-instance learner for tissue image classification

The multi-instance learning method with certainty value-based pooling functions addresses the limitations of existing image classification techniques in digital pathology, enhancing accuracy and interpretability for whole-slide images by incorporating model uncertainty and predictive feature identification.

JP7583041B2Active Publication Date: 2024-11-13F HOFFMANN LA ROCHE & CO AG
View PDF 1 Cites 0 Cited by

Patent Information

Application Number
JP2022523835
Authority / Receiving Office
JP · JP
Patent Type
Patents
Current Assignee / Owner
Priority Date
2019-11-22
Filing Date
2020-11-20
Publication Date
2024-11-13
Estimated Expiration
2040-11-20

AI Technical Summary

Technical Problem

Existing machine learning techniques for image classification in digital pathology, such as breast cancer detection, struggle to classify whole-slide images effectively and lack interpretability, often functioning as 'black boxes' that make it difficult for physicians to understand the decision-making process.

Method used

A multi-instance learning (MIL) approach with certainty value-based pooling functions is employed to classify tissue images, using certainty values to aggregate feature vectors and predicted values, enhancing model interpretability and accuracy, especially for whole-slide images.

Benefits of technology

The MIL approach improves classification accuracy and interpretability by considering model uncertainty, allowing for more robust classification of tissue images, even with low evidence rates, and enabling the identification of predictive features without requiring explicit medical knowledge.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 0007583041000013
    Figure 0007583041000013
  • Figure 0007583041000014
    Figure 0007583041000014
  • Figure 0007583041000015
    Figure 0007583041000015
Patent Text Reader

Abstract

The present invention relates to a method for classifying tissue images, the method comprising: receiving (102) a plurality of digital tissue images; Dividing each received image into a set of image tiles (104); For each of the tiles, extracting a feature vector from the tile (106); providing a multi-instance learning (MIL) program (108) configured to use the model to classify any input image as a member of one of at least two different classes based on feature vectors extracted from all tiles of the image; calculating (110) for each tile a certainty value indicating the model's certainty regarding the contribution of the tile's feature vector to the classification of the image; For each image, using a MIL program a certainty value-based pooling function to aggregate the image feature vector or predicted values ​​calculated from the image feature vector into an aggregate predicted value as a function of the certainty values ​​of the tiles of the image (114); classifying (116) each of the images as a member of one of the classes based on the aggregate prediction value; Includes.
Need to check novelty before this filing date? Find Prior Art

Description

[Technical field]

[0001] FIELD OF THEINVENTION The present invention relates to the field of digital pathology, in particular to the field of image analysis. [Background technology]

[0002] Background and Related Art There are several known image classification methods that can be used to classify digital pathology images into different categories, such as "healthy tissue" or "cancerous tissue." For example, Sertan Kaymaka et al., in "Breast cancer image classification using artificial neural networks," Procedia Computer Science, Volume 120, 2017, pages 126-131 (Non-Patent Document 1), describe an automatic classification method for breast cancer diagnostic images using a backpropagation neural network (BPPN).

[0003] However, applicants have observed that various machine learning techniques that provide good results in early detection of cancer-related nodes in mammography images are unable to classify images of other types of tissue sections, particularly whole slide images.

[0004] A further problem associated with the use of existing machine learning approaches for image classification is that the trained machine learning programs often function like black boxes, which can be frustrating for doctors and patients who must rely fully or partially on this "black box" to decide whether it makes sense to prescribe a potentially effective, but side-effect-ridden, drug to a particular patient without being able to verbalize the underlying "decision logic."

[0005] MAXIMILIAN ILSE et al.: "Attention-based Deep Multiple Instance Learning", ARXIV.ORG, CORNELL UNIVERSITY LIBRARY, 201 OLIN LIBRARY CORNELL UNIVERSITY ITHACA, NY 14853, February 13, 2018, XP081235680 (Non-Patent Document 2) describes the application of an attention-based multi-instance learner (MIL) to histopathology datasets.

[0006] The anonymous publication "DEEP MULTIPLE INSTANCE LEARNING WITH GAUSSIAN WEIGHTING", ICLR 2020 Conference Blind Submission, September 25, 2019 (2019-09-25), pages 1-10, XP055698116, retrieved from the Internet on May 25, 2020 from URL: https: / / openreview.netiattachment?id=Bklrea4KwS&name=original.pdf (Non-Patent Document 3) describes a deep multiple instance learning (MIL) method that is trained end-to-end to perform classification from weakly supervised learning. The MIL method is implemented as a two-stream neural network specialized for the task of instance classification and weighting utilizing Gaussian radius basis functions that normalize instance weights by comparing instances locally within a bag and globally across multiple bags. [Prior art documents] [Non-patent literature]

[0007] [Non-Patent Document 1] Sertan Kaymaka et al., “Breast cancer image classification using artificial neural networks”, Procedia Computer Science, Volume 120, 2017, pages 126-131 [Non-Patent Document 2] MAXIMILIAN ILSE et al.: "Attention-based Deep Multiple Instance Learning", ARXIV.ORG, CORNELL UNIVERSITY LIBRARY, 201 OLIN LIBRARY CORNELL UNIVERSITY ITHACA, NY 14853, February 13, 2018, XP081235680 [Non-Patent Document 3] "DEEP MULTIPLE INSTANCE LEARNING WITH GAUSSIAN WEIGHTING", ICLR 2020 Conference Blind Submission, September 25, 2019 (2019-09-25), pages 1-10, XP055698116, URL: https: / / openreview.netiattachment?id=Bklrea4KwS&name=original.pdf Summary of the Invention

[0008] overview It is an object of the present invention to provide an improved method for classifying tissue images and a corresponding image analysis system as claimed in the independent claims. Embodiments of the invention are set out in the dependent claims. The embodiments of the invention can be freely combined with one another if they are not mutually exclusive. In one aspect, the present invention relates to a method for classifying tissue images, the method comprising: - receiving, by an image analysis system, a plurality of digital images, each of the digital images representing a tissue sample from a patient; dividing each received image into a set of image tiles by an image analysis system; - calculating, for each of the tiles, a feature vector comprising image features selectively extracted from the tile by an image analysis system; - providing a multi-instance learning (MIL) program configured to use a model for classifying any input image as a member of one of at least two different classes based on feature vectors extracted from all tiles of said input image; - calculating, for each of the tiles, a certainty value (called herein "c" according to an embodiment of the invention), the certainty value indicating the certainty value of the model regarding the contribution of the feature vector of the tile to the classification of the image from which the tile is derived; -For each image, using a certainty value based pooling function for aggregating feature vectors extracted from the image into a global feature vector as a function of the certainty values ​​of tiles of the image, by the MIL program, and calculating an aggregate prediction value (herein called "ah" according to an embodiment of the present invention) from the global feature vector; or calculating, by a MIL program, a prediction value from each of the feature vectors of the image, and using, by the MIL program, a certainty value-based pooling function to aggregate the prediction values ​​of the image into an aggregate prediction value (referred to herein as "ah" according to an embodiment of the present invention) as a function of the certainty values ​​of the tiles of the image; - classifying, by a MIL program, each of the images as a member of one of at least two different classes based on the aggregate predictive value.

[0009] These features can be beneficial for a number of reasons.

[0010] A multi-instance learning (MIL) program is a form of weakly supervised learning program configured to learn from a training set, where training instances are arranged into sets called bags, and the labels of individual instances in the bags are not known, but a label is provided for the entire bag. Thus, MIL programs require only weakly annotated training data. This type of data is particularly common in medical imaging, since annotation of individual image regions to provide richly annotated training data is very time-consuming and therefore expensive. Furthermore, tissue structures that signify (have high predictive value) that a digital image is a member of a particular class (e.g., images showing healthy tissue / images showing primary tumors / images showing metastases) may not be known or recognizable to a pathologist. Thus, using MIL programs to classify digital tissue images can have the advantage that weakly annotated training data is sufficient to train a MIL program that can accurately classify digital tissue images. Furthermore, a trained MIL program can accurately classify digital tissue images even when a human annotator, e.g., a pathologist, does not know the tissue structure that is highly predictive of the tissue's class membership and is therefore unable to select training images with an unbiased ratio of tissue regions with and without this tissue structure.

[0011] Furthermore, the use of the MIL program with a certainty value based pooling function incorporates model uncertainty into the classification, which Applicant has observed to significantly improve classification accuracy, especially in the area of ​​tissue slide image analysis, especially when the tissue slide images are whole slide images.

[0012] When the MIL program is used to solve a computational pathology problem, whole slide images (WSIs) used as training images are given a global label (e.g., indicating whether tumor cells are present in the biopsy). Multiple instances are then extracted from the training WSI by sampling image tiles from the training WSI and grouped into bags, where each bag contains tiles extracted from a particular training WSI and has the global label for that slide image.

[0013] In many cases, for example when a tumor is localized in a small portion of a biopsy, only a small proportion of the instances (tiles) contain evidence of WSI labeling. In addition, the size of the bag (number of tiles per training WSI) can be very large due to the large size of tissues in the WSI at full resolution (on the order of thousands or more instances). These factors make for a challenging MIL setting. As the bag gets larger, the probability of the bag being misclassified increases, since a larger population of negative instances in the bag have a greater chance of finding evidence of the positive class. This is amplified by the unstable nature of deep learning models, where small changes in the input image can cause very different outputs. For this reason, a large bag containing many visually similar looking instances can result in very different feature vectors and their respective predictions and predicted values. Applicant has observed that this problem is significantly improved by considering not only the feature values ​​of the tile (or predicted values ​​derived from the feature vectors) but also certainty values, since the model's uncertainty regarding the tissue texture and image features depicted in a particular tile is taken into account.

[0014] According to an embodiment, the certainty value based pooling function is a certainty value based max pooling, average pooling or attention pooling function.

[0015] Pooling functions are a key element in MIL. Pooling functions determine how predictions of instances, i.e. tiles of an image, of the MIL model are combined to form a bag output, i.e., classification results for the image. There are several pooling functions, e.g., max pooling, average pooling, and attention pooling. However, applicants have observed that for low evidence rate bags (e.g., a small number of positive instances compared to the total number of instances), when max pooling is used, a single false positive instance prediction corrupts the resulting bag prediction, producing a false positive result. On the other hand, when average pooling is used, a large population of negative instances in the bag will obscure the positive instances, producing a false negative bag prediction.

[0016] In the case of learned attention MIL, attention learning is highly adversely affected in the case of low evidence rates. An additional challenge with large bag poses is the interpretability of the model by selecting key instances. Larger bags present a higher probability of instability potentially leading to errors in key instance selection. Therefore, the use of a certainty value-based pooling function is particularly advantageous in the context of digital tissue images, especially whole slide images.

[0017] The use of a certainty-value based pooling function addresses the shortcomings of current pooling functions and the underperformance of MIL in the case of low-evidence bags, which is common in the case of (whole-slide) tissue images. For example, the model uncertainty for every instance is taken into account by weighting the feature vector (or the prediction derived from it) with the certainty value calculated for this instance.

[0018] Classification of digital tissue images can be used to evaluate the possibility of successfully treating a patient suffering from a disease with a specific drug. For example, some drugs used in the course of immunotherapy in cancer patients are only effective if specific immune cells are found at a specific distance from the cancer cells. In this case, an attempt is made to automatically recognize those objects, i.e. specific cell types or specific lower and higher cellular structures, in the tissue images in order to be able to make a statement about the presence of the disease and / or its recommended treatment. Embodiments of the present invention can have the advantage that they utilize MIL programs and therefore do not require that the relationship between specific tissue structures and specific diseases or their treatment options be explicitly known. By training and using trained MIL programs, it is possible to implicitly detect unknown predictive features regarding a specific disease and / or its treatment. Thus, embodiments of the present invention are not limited to the medical knowledge available at a particular time.

[0019] In a further beneficial aspect, using a MIL program that treats image tiles as cases is particularly suited to predicting patient-related features in the context of the entire slide tissue sample image. This is because the entire slide tissue sample often covers many different tissue regions, only some of which may have any predictive value. For example, a micrometastasis may be only a few millimeters in diameter, while the slide and the respective entire slide image may be several centimeters long. Although the entire image is labeled according to empirical observations for the patient from which the sample was derived, with a particular label such as "response to drug D=true", the tissue region around the micrometastasis, which contains many immune cells and is predictive for a positive response, may also cover only a few millimeters. Thus, the majority of tiles do not contain any tissue regions that are predictive in terms of the label per image, typically per patient. The MIL program is particularly suited to identifying predictive features based on a bag of data instances, where the majority of the instances are presumed not to be of any predictive value.

[0020] MIL is configured to treat all tiles derived from the digital image as members of the same bag of tiles.

[0021] According to an embodiment, the method further includes outputting the classification results to a user via the GUI. Additionally or alternatively, the method includes outputting the classification results to another application program. For example, the MIL program can be part of or interoperable with an image classification application program configured to generate a GUI that displays the results of the classification to a user. For example, each of the received digital images can be displayed in the GUI in association with a label indicating the class that the respective digital image belongs to according to the classification results generated by the MIL program.

[0022] According to an embodiment, the MIL program is a binary MIL program. The at least two classes consist of a first class called the "positive class" and a second class called the "negative class". Any one of the images is classified into the "positive class" if the MIL model predicts that for at least one of the tiles of the image, the feature vector of the tile contains evidence of the "positive class". Any one of the images is classified into the "negative class" if the MIL model predicts that for all tiles of the image, their respective feature vectors do not contain evidence of the "positive class". For example, the question of whether a sufficient number of tiles contain sufficient evidence of the "positive class" can include using a certainty value-based pooling function to determine whether the aggregate prediction value exceeds a threshold.

[0023] In some embodiments, the feature vector computed for a tile can include one or more "positive feature values" and one or more "negative feature values." A "negative feature value" is a value of a feature that is evidence of a "negative class." A "positive feature value" is a value of a feature that is evidence of a "positive class." In this case, the MIL program is configured to consider both negative and positive feature values ​​for classifying an image.

[0024] The predicted membership of an image in a particular class can be implemented as assigning a class label to the image that indicates the class membership. The "class label" assigned to a particular digital image can also be called a "bag label" because the image represents a "bag" and the tiles generated from the image represent "instances" of the bag.

[0025] For example, the MIL program can be a binary MIL program in which a binary label is assigned to every bag. The MIL program is configured to learn and predict that a particular digital image has a "positive" bag label if at least one of the instances (tiles) of the image contains evidence of the label. If all instances do not contain evidence of the "positive" class label, the image is predicted to have a "negative" bag label.

[0026] More formally, every bag (image) is a set of instances {x 1 ,...,x K}, where K is the size of the bag, i.e., the number of tiles generated from a digital image. K can vary between bags (incoming digital images). Each bag (image) is associated with a binary class label Y ∈ {0,1}. Every instance (tile) j is also associated with a label y j ∈{0,1}, where instance labels (tile labels) are assumed to be hidden. "0" represents a negative instance or bag label and "1" represents a positive instance or bag label. MIL programs are configured to predict a binary bag label (class membership) Y of a bag (image) by applying a pooling function to all instance labels to compute / predict the bag label. For example, a state-of-the-art MIL program using a max pooling function computes the binary bag label Y according to: Y={0,iffΣy K = 0 otherwise 1}. However, the MIL program according to an embodiment of the present invention uses a new pooling function that takes into account the uncertainty in the model.

[0027] A MIL program contains a pooling function that is applied to aggregate the instance predictions (individual tile-based predictions), h, to create a bag (image) prediction (classification result).

[0028] Approach I based on tile-based predictors According to an embodiment, the MIL program uses a certainty-value-based pooling function to aggregate the predicted values ​​calculated for tiles of an image into an aggregate predicted value. The method includes the MIL program calculating one of the predicted values ​​(referred to herein as "h" according to an embodiment of the invention) for each of the tiles. Each predicted value is calculated as a function of the feature vector extracted from the tile. A predicted value is a data value indicating the contribution of the tile's feature vector to the classification of the image from which the tile was derived. When calculating the predicted value for a particular tile, the MIL program preferably takes into account only the feature vector of the tile from which the feature vector was derived (and not the feature vectors of other tiles).

[0029] According to an embodiment, the certainty value-based pooling function is a certainty value-based maximum pooling function. The use of the certainty value-based pooling function includes, for each of the images, a sub-method a) or b, where the sub-method a) or b, respectively, includes: - a1) weighting the prediction value (e.g., called "h") of each tile by the certainty value (e.g., called "c") calculated for this tile, thereby obtaining a weighted prediction value (e.g., called "wh"); for example, the weighting may include multiplying the certainty value (c) calculated for one of the tiles by the prediction value (h) calculated for that one tile to calculate the weighted prediction value (wh) of the tile, e.g., for a particular image I divided into K tiles, m For any one of the K tiles, the weighted prediction value wh m_k can be calculated as follows: m_k =h m_kX c m_k , -a2) The maximum of all weighted predictions calculated for all tiles of the image (e.g. "wh MAX For example, the maximum value is calculated according to: TIFF0007583041000001.tif5128 and -a3) Maximum weighted predicted value (whMAX (I m )) as the aggregate predictor; or -b) Maximum certainty value (c max ) as the aggregate prediction.

[0030] If the prediction (h) of the tile with the highest certainty value (c) is used as the aggregate prediction, then the certainty value c is calculated only to select the prediction h. In this case, the selected prediction h is considered to be "implicitly" weighted by the certainty value, since it is selected as the aggregate prediction depending on the certainty values ​​calculated for the tiles of the image.

[0031] The aggregate predictions are then used to classify the input image. MAX If x exceeds a threshold determined during training, the digital image can be classified as an "image indicative of a tumor." Otherwise, the image is classified as an "image indicative of healthy tissue."

[0032] Using a certainty value based max-pooling function can have the advantage that the MIL program is more robust to situations during testing or training when a large population of negative instances (tiles with "negative class" tile labels) in a bag (image) can potentially obscure positive instances (tiles with "positive class" labels) and potentially form false negative bag class predictions.

[0033] According to an embodiment, providing the MIL program includes training the MIL program on a training image set, whereby during the training phase, a certainty value-based max pooling function is used as the pooling function. This can have the advantage that the predictive model generated during the training of the MIL strongly reflects the tissue patterns shown in the tiles with the feature vectors that have the highest predictive power for the bag label. The model is not adversely affected by tissue regions / tiles that are irrelevant for the label. However, the max operation ignores all information contained in all tiles except the highest scored tile. Therefore, the predictive power of tiles / tissue patterns that may be relevant may be missed.

[0034] According to an embodiment, providing the MIL program includes training the MIL program on a training image set, whereby during the training phase, a certainty value-based average pooling function is used as a pooling function. This can be beneficial because the predictive model generated when training the MIL program takes into account the tissue patterns depicted in all tiles. However, the consideration of tissue patterns that are actually unrelated to the occurrence of a particular label and each tile may result in degradation and a decrease in the predictive accuracy of the trained MIL.

[0035] According to an embodiment, the certainty value based pooling function is a certainty value based average pooling function. The use of the certainty value based pooling function includes, for each of the images: - weighting the prediction value of each tile (e.g., called "h") by the certainty value (e.g., called "c") calculated for this tile, thereby obtaining a weighted prediction value (e.g., called "wh"); for example, weighting includes multiplying the certainty value (c) calculated for one of the tiles by the prediction value (h) calculated for that one tile to calculate the weighted prediction value (wh) of the tile; for example, for a particular image I divided into K tiles, m For any one of the K tiles, image I m The weighted prediction value wh generated by the model m_k can be calculated as follows: m_k =h m_kX c m_k ; - The average of all weighted predictions wh calculated for all tiles of the image (e.g. "wh MEAN For example, the average may be calculated according to: According to another embodiment, the average is calculated according to: TIFF0007583041000003.tif5128 - and the average weighted predicted value (wh MEAN (I m )) as the aggregate predictor.

[0036] Using a certainty value-based average pooling function can have the advantage that the MIL program is more robust to false positive predictions that an image is a member of the "positive class" in low evidence rate bag situations than a max operator-based pooling function. A "low evidence rate bag" is a bag of instances where the number of positive instances (tiles with a "positive class" label) is very low compared to the total number of instances (tiles). When max pooling is used, a single false positive instance prediction destroys the resulting bag prediction, creating a false positive classification result.

[0037] Global aggregate feature vector based approach II According to an alternative embodiment, the MIL program uses a certainty-value-based pooling function to aggregate feature vectors extracted from tiles of an image into a global ("aggregate") feature vector that is again used to calculate an aggregate prediction value. The method includes: - applying a pooling function to the feature vectors and certainty values ​​calculated for the tiles of the image in order to calculate a global feature vector of the image, the calculation of the global feature vector taking into account the feature vectors and certainty values ​​of the tiles; and - Using a global feature vector to calculate aggregate predictions.

[0038] According to an embodiment, the certainty value based pooling function is a certainty value based max pooling function.

[0039] According to one embodiment, the use of the certainty value based max pooling function comprises, for each of the images, sub-methods c) or d), respectively: - c1) weighting each feature vector of a tile (e.g. called "fw") by the certainty value (e.g. called "c") calculated for this tile, thereby obtaining a weighted feature vector (e.g. called "wfv"); for example, the features of the feature vector may consist only of numerical feature vectors; the weighting consists in multiplying the certainty value (c) calculated for one of the tiles by each feature value in the feature vector (fv) calculated for one tile to obtain a weighted feature vector (fv) for the tile; -c2) The maximum of all weighted feature vectors (wfv) calculated for all tiles of the image max ) to identify; or -d) Maximum certainty value (c max ) as a global feature vector.

[0040] For example, an embodiment according to sub-method d) can be used when the feature vector contains ordinal values ​​that cannot be weighted by simple multiplication.

[0041] According to another embodiment, the certainty value based pooling function is a certainty value based average pooling function. Use of the certainty value based pooling function includes, for each of the images: - weighting the feature vector (e.g. called "h") extracted from each of the tiles by a certainty value (e.g. called "c") calculated for this tile, thereby obtaining a weighted feature vector (e.g. called "wfv"); for example, each feature vector may include multiple numerical feature values, each numerical feature value of the feature vector being multiplied by a certainty value to obtain a weighted feature vector value stored in the weighted feature vector for the particular tile; - calculating an average weighted feature vector from all weighted feature vectors (wfv) calculated for an image; for example, the feature values ​​of the average weighted feature vector can be calculated as the average of all weighted feature values ​​stored in the weighted feature vectors of images in the same vector position; and using the average weighted feature vector calculated for the image as a global feature vector for the image used to calculate the aggregate predicted value.

[0042] Further embodiments (approaches I and II) According to another embodiment, the method includes normalizing the certainty values ​​calculated for the images by applying, for each of the images, a softmax function to the certainty values. The certainty value-based pooling function is configured to aggregate the prediction values ​​as a function of the normalized certainty values ​​into an aggregate prediction value, or to aggregate the feature vectors as a function of the normalized certainty values ​​into a global feature vector, the global feature value being used to calculate the aggregate prediction value.

[0043] The softmax function, also known as softargmax or normalized exponential function, is a function that takes an array of K real numbers, e.g. an image I m The K certainty values ​​c k Softmax is a function that takes as input K as a normalization function and normalizes it into a probability distribution consisting of K probabilities proportional to the exponent of the number of inputs. That is, before applying softmax, some array components may be negative, or greater than 1, and may not sum to 1; after applying softmax, each component lies in the interval (0,1), and the components sum to 1, so they can be interpreted as a probability. Furthermore, larger input components correspond to larger probabilities. Softmax maps the unnormalized output of the network into a probability distribution over the predicted output classes or ranges.

[0044] For example, a softmax function can be used to normalize the certainty values ​​to a numerical range of values ​​from 0 to 1. Using a softmax function to normalize the certainty values ​​can have the advantage that the comparability of the certainty values ​​obtained for different tiles and / or different images is increased.

[0045] According to an embodiment, the MIL program is configured to include and perform a certainty value based pooling function instead of one or more of the following (traditional) pooling functions: A max-pooling function configured to identify and return the maximum of all predictions h computed from all feature vectors extracted from a tile of images; for example, for a given image I m The conventional aggregate prediction for Can be calculated as TIFF0007583041000004.tif7128; An average pooling function configured to identify and return the average of all predictions h computed from all feature vectors extracted from a tile of images; for example, for a given image I m The conventional aggregate prediction for Can be calculated as TIFF0007583041000005.tif7128; An attention pooling function that identifies and returns a prediction value h of one of all feature vectors extracted from a tile of images that the attention learning technique identifies as having the most predictive power for the class membership of the image from which the tile was derived; for example, for a particular image I m The conventional aggregate prediction value for It can be calculated as TIFF0007583041000006.tif7128.

[0046] Thus, according to an embodiment of the present invention, the new certainty-based pooling function replaces the conventional pooling function used in conventional MIL programs (typically or selectively in the training or testing phase). The MIL program according to an embodiment of the present invention does not include or use any of the above three conventional pooling functions for determining bag labels / performing image classification tasks (typically or selectively in the training or testing phase).

[0047] According to an embodiment, the MIL program is a neural network, and the certainty value is calculated using a dropout technique during training and / or testing of the model of the neural network.

[0048] Dropout is a technique that randomly switches off some neurons from a fully connected layer. Typically, dropout is applied (only) during training. Dropout forces a fully connected layer to learn the same concept in a different way. Dropout means that a certain percentage of neurons in a certain layer are randomly deactivated ("dropped out"). This improves the generalization ability of the trained model, since the layer to which dropout is applied is forced to learn the same "concept" with a different set of interconnected neurons. Dropout is therefore a technique that can be used to avoid overfitting of neural networks during training. In general, the higher the learning ability of a model / neural network (more layers, or more neurons), the more prone the model / NN is to overfit.

[0049] Using the dropout technique to calculate the model certainty of each tile prediction can have the advantage that existing program libraries and software tools can be used to calculate the model uncertainty, since many neural network architectures and programs already include one or more dropout layers. Furthermore, applicants have observed that neural networks used in the context of digital pathology often face problems with overfitting, since neural networks used to perform image analysis and classification tasks often have many layers and the training datasets are often limited in size.

[0050] According to an embodiment, applying dropout during the training phase means creating many different dropout layers, each containing a randomly selected subset of fully connected layer nodes, whereby each dropout layer acts as a mask (containing "0" and "1" nodes). The masks are created during forward propagation and applied to the layer outputs respectively during training and cached for future use in backpropagation. The dropout masks applied to the layers are stored so that activated neurons can be identified during the backpropagation step. Now, once the identified neurons are selected, the neuron's output is backpropagated. Typically, dropout layers are created and used only during the training phase and are stored in the form of deactivated additional layers in the trained network.

[0051] According to some embodiments, the dropout layer is only used during training.

[0052] This can have the advantage that the trained MIL program has already learned during the training phase to evaluate the variability of the predictions computed for the feature vectors of a particular tile using many different network architectures that use dropout masks. If the training images are similar to the tissue images used during testing, the model uncertainty / variability of the computed tile-based predictions is intrinsically encoded into the trained ML program and accurately reflects the model uncertainty for the various tissue structures shown in the test images.

[0053] Some embodiments of the present invention use fully connected layer dropout, while other embodiments also use dropout after a max pooling layer, thereby generating some kind of image noise enhancement.

[0054] During the testing phase, dropout layers are typically deactivated in state-of-the-art machine learning programs. However, according to some embodiments, dropout techniques are used during model training and testing. For example, the creation or reactivation and application of dropout layers allows the same concept to be learned (during training) or predicted (during testing) in many different ways, thereby allowing the model uncertainty and the diversity of predictions generated by the model for a given input to be evaluated. According to embodiments, the certainty value is calculated as Monte Carlo dropout (MC dropout).

[0055] The key idea behind Monte Carlo Dropout (MC Dropout) is to evaluate model uncertainty using dropout. MC Dropout calculates model uncertainty (and therefore, implicitly, also model certainty) using a dropout technique, i.e., by randomly using different sub-networks of a neural network architecture to obtain multiple different results from the network for the same prediction task, and evaluates "certainty" as the "consistency" of the results. The dropout process is similar to sampling neurons, hence MC refers to the Monte Carlo method.

[0056] For example, according to some embodiments of the invention, at test time the same input is presented to the network multiple times, e.g., hundreds of times, with random dropout, using a different dropout mask each time. The average of all the multiple predictions obtained for each tile is then calculated to generate a prediction interval or another measure of prediction variability / model uncertainty that covers all these predictions.

[0057] Applying dropout at test time may have the drawback of reducing the overall model accuracy. However, in the context of tissue image analysis, it has been observed that this drawback is in fact beneficial, since this step imposes uncertainty on the model, and will and can reduce overfitting. For example, if the input data is far from the data on which the model was trained, the variability of predictions obtained at test time for the same input tile using many different dropout masks indicates that the model is unclear how to interpret this type of tile. In the field of tissue image analysis, a great variety of healthy and disease-induced tissue structures exists. Further variability is induced by the fact that there are many different staining protocols, the presence of many different stains, and even the characteristics of the individual person performing the staining protocol can affect the stained tissue section and therefore the digital tissue image. Therefore, it is highly unlikely that a training dataset will cover all possible variations and combinations of healthy and diseased tissue structures and different staining methods. Thus, the use of MC dropout provides a MIL program that, at test time, can automatically assess, for each of the tiles individually, the uncertainty of a model configured to predict the correct class of an image based on the features of this particular tile.

[0058] According to some embodiments, MC Dropout Mean-STD (MC Dropout Mean-Standard Deviation), which involves measuring the variance of the model over multiple forward pass runs, is used to calculate the model's confidence in a particular classification task and is used to calculate a confidence value for each of the tiles. Tiles whose feature vectors do not contain sufficient and / or relevant information for the model to confidently and accurately determine the bag label will generate low prediction values ​​during training. An example for calculating MC Dropout Mean-STD is described in detail in Yarin Gal et al., "Deep Bayesian Active Learning with Image Data," March 2017, arXiv:1703.02910v1.

[0059] According to an embodiment, the certainty value based pooling function of the MIL program is configured to calculate, for each tile of the image, a prediction value wh weighted by the MC dropout mean-STD as follows:

[0060] Image I m For each of the K tiles in, we give a certainty value (also called the “instance certainty”) c k_m is calculated as the inverse of the mean-STD based on the MC dropout calculated by the model in the MIL program after the softmax or sigmoid layer at test time. The formula for the confidence of all instances (tiles) k, c k_m teeth, TIFF0007583041000007.tif7128, where σ(x k ) is the MC dropout mean-STD for instance k, and ε is a small number that prevents division by 0.

[0061] The certainty value c of every tile in the image k_mcan be aggregated using a maximum or average-based certainty value-based pooling function, as described above for embodiments of the present invention. If a certainty value-based maximum pooling function is used, the instance (tile) with the highest predicted value h produced as output by the model of the MIL program weighted by its certainty value c is selected. If a certainty value-based average pooling function is used, according to an embodiment of the present invention, the certainty value c of each of the K tiles is passed through a softmax function (which can be implemented as a layer of a network) to obtain an average weight that sums to 1. Then, a weighted average wh of the instance predictions h using the certainty value c as weight is calculated. Thus, whk is the output of the predictive network for instance (tile) k.

[0062] An embodiment of the present invention provides a new certainty-value-based pooling function that aggregates bag instances using certainty values ​​assigned to each individual instance. The certainty-value-based pooling function can be applied during training, but can also be used to improve pre-trained MIL programs that are trained with dropout at test time.

[0063] According to an embodiment, a certainty value based pooling function (which can use dropout techniques to calculate the model uncertainty) is used during testing but not during model training.

[0064] This means that, according to embodiments of the present invention, image classification methods can be applied "off the shelf" to pre-trained MIL models to significantly improve their performance, but during training, the certainty-based pooling function may not have been used. This can have the advantage that even if a MIL program is trained without using a certainty-value-based pooling function (with or without the dropout technique), image analysis methods can be used to classify images, whereby the existing dropout layers (but initially deactivated for test-time use) are applied at test time to generate various predictions for specific image tiles, thereby assessing the certainty of the model with respect to the tile-related predictions.

[0065] For example, the MIL program trained on the training dataset may not include any dropout layers at training time, and the pooling function used at training time does not consider the uncertainty of the model. For example, the pooling function used by the MIL program at training time can be a traditional max pooling function or an average pooling function. Calculating the certainty value for any one of the tiles at test time includes adding one or more dropout layers to the trained MIL program at test time. The added dropout layers can be used at test time to calculate a certainty value for each input tile that reflects the uncertainty of the trained model.

[0066] According to an embodiment, the neural network includes one or more deactivated dropout layers, the deactivated dropout layers being dropout layers that are activated during training and deactivated upon completion of training, and the calculation of the certainty value for any one of the tiles during testing includes reactivating the one or more dropout layers during testing.

[0067] According to an embodiment, the calculation of the certainty value for any one of the tiles at test time includes, after adding or reactivating one or more dropout layers at test time: For each tile, a predicted value h is calculated based on the feature vector extracted from the tile. d multiple times to obtain the predicted value h d Each time,x,i,is computed, a different subset of the nodes of the network is,reactivated or dropped by an additional dropout layer to form,x,i,. - For each tile, multiple predicted values ​​h calculated for the tile d and calculating a certainty value c for the tile as a function of variability of the tile, where the greater the variability, the lower the certainty value c.

[0068] This can have the advantage that if the network's existing dropout layers are deactivated after the training phase is completed, or if they are used with a significantly different dataset during the training phase, or if the neural network being used as the MIL program does not include a dropout layer at test time, reactivating existing layers or adding a dropout layer at test time can make it possible to generate different predictions for particular image tiles at test time, thereby assessing the robustness of the model with respect to tile-related predictions, even if the network architecture of the trained model does not initially allow for this variance to be calculated.

[0069] For example, the MIL program may be trained on digital tissue images that are significantly different from the tissue images used at test time. For example, the training and test images may be from patients with different types of cancer, or may be similar but stained with different stains, or may depict the same type of cancer cells but different tissues. In these cases, the trained model may not be able to provide highly accurate predictions for the tissue images and image tiles analyzed at test time. However, by applying dropout techniques at test time, the accuracy of the predictions h generated by the trained model on the feature vectors extracted from the tiles can be automatically evaluated and taken into account to classify the tissue images provided as input at test time.

[0070] According to an embodiment, the received digital images include digital images of tissue samples in which pixel intensity values ​​correlate with the amount of non-biomarker specific staining, in particular hematoxylin or H&E staining.

[0071] For example, each bag of tiles can represent a respective patient whose response to a particular drug is known. The instances contained in this patient-specific bag are tiles derived from one or more images of respective tissue samples of this particular patient, where the tissue samples have been stained with a non-biomarker-specific stain, such as H&E. All tissue images of this patient, and therefore all tiles derived from them, are assigned the label "Response to Drug D = True Patient". H&E stained tissue images represent the most common form of stained tissue images, and this type of staining alone already reveals a lot of data that can be used to predict the value of patient-related attributes, such as a particular tumor subtype or stage. Furthermore, many hospitals contain large databases of H&E stained tissue images derived from patients they have treated over the years. Typically, hospitals also have data on whether a particular patient responded to a particular treatment and / or how fast or severe the disease progression is. Thus, a large corpus of training images is available, where each outcome can be labeled (e.g., efficacy of treatment with a particular drug (yes / no), progression-free survival >1 year, progression-free survival >2 years, etc.).

[0072] According to an embodiment, the received digital images include digital images of tissue samples whose pixel intensity values ​​correlate with the amount of biomarker-specific dyes. The biomarker-specific dyes are stains adapted to selectively stain the biomarkers contained in the tissue samples. For example, the biomarkers can be specific proteins such as HER-2, p53, CD3, CD8, etc. The biomarker-specific dyes can be stains for bright field or fluorescence microscopy that are bound to antibodies that selectively bind to the biomarkers.

[0073] Additionally or alternatively, the received digital images include digital images of the tissue sample in which pixel intensity values ​​correlate with amounts of a biomarker-specific staining agent, the biomarker-specific staining agent being adapted to selectively stain a biomarker contained in the tissue sample.

[0074] For example, each bag of tiles can represent a respective patient whose response to a particular drug is known. The instances contained in this patient-specific bag are tiles derived from one or more images of respective tissue samples of this particular patient. The one or more tissue samples are stained with one or more biomarker-specific dyes. For example, a tile can be derived from one, two, or three tissue images all depicting adjacent tissue slides of the same patient stained with a HER2-specific dye. According to another example, a tile can be derived from a first tissue image depicting a first tissue sample stained with a HER2-specific dye, a second tissue image depicting a second tissue sample stained with a p53-specific dye, and a third tissue image depicting a third tissue sample stained with a FAP-specific dye. The first, second, and third tissue samples are derived from the same patient. For example, they can be adjacent tissue sample slices. These three tissue images depict three different biomarkers, but all tissue images are derived from the same patient. Thus, all tiles derived from them are assigned the label "Response to Drug D = True Patient". Having the MIL program train on image tiles of digital images whose pixel intensity values ​​correlate with the amount of biomarker-specific dye can have the advantage that identifying the presence and location of one or more specific biomarkers in tissue can highly reveal specific predictive information with respect to specific diseases and disease sub-forms. Predictive information can include observed positive and negative correlations with the presence of two or more biomarkers. For example, the recommended treatment schemes and predictions of some diseases, such as lung or colon cancer, are observed to be highly dependent on the mutational signature and expression profile of the cancer. Sometimes the expression of a single marker does not have predictive power by itself, but the combined expression of multiple biomarkers and / or the absence of certain additional biomarkers can have high predictive power with respect to certain patient-related attribute values.

[0075] According to an embodiment, the received digital images include a combination of digital images of tissue samples whose pixel intensity values ​​correlate with the amount of a first biomarker-specific dye and digital images of tissue samples whose pixel intensity values ​​correlate with the amount of a non-biomarker-specific dye. The biomarker-specific dye is a staining agent adapted to selectively stain a biomarker contained in the tissue sample. The MIL program is trained to classify all digital images showing the same tissue sample and / or showing adjacent tissue samples from the same patient into the same class.

[0076] This approach can have the advantage that identifying the presence and location of one or more specific biomarkers in tissue in combination with an informative tissue signature revealed by H&E staining can provide highly specific predictive information with respect to specific diseases and disease sub-forms. The predictive information can include observed positive and negative correlations with the presence of two or more biomarkers and / or with tissue signatures visually revealed by H&E staining.

[0077] According to an embodiment, providing the MIL program includes training a model of the MIL program. - providing a set of digital training images of tissue samples, each digital training image being assigned a class label indicative of one of at least two classes; - dividing each training image into training image tiles, each training tile being assigned the same class label as the digital training image from which the training tile was derived; - calculating, for each of the tiles, a training feature vector comprising image features selectively extracted from the tile by the image analysis system; and / or - iteratively adapting a model of the MIL program to minimize an error of a loss function, the error of the loss function indicating the difference between the predicted class labels of the training tiles and the class labels actually assigned to the training tiles, the predicted class labels being calculated by the model based on the feature vectors of the training tiles.

[0078] According to an embodiment, the method comprises the steps of: for each received digital image: - weighting the prediction value h of each tile with the certainty value c calculated for this tile, thereby obtaining a weighted prediction value wh; - Identifying one of the tiles of the image for which the highest weighted prediction value wh is calculated by the MIL program; - calculating, for each of the other tiles of the image, a relevance indicator by comparing the weighted prediction value wh of the other tile with the highest weighted prediction value, the relevance indicator being a number that is negatively correlated with the difference between the compared weighted prediction values; - calculating a relevance heatmap of the image as a function of the relevance indicators, the pixel colors and / or pixel intensities of the relevance heatmap being indicative of the relevance indicators calculated for tiles in the image; - displaying a relevance heatmap on the GUI.

[0079] This can be advantageous because it can enable a pathologist to identify the tiles and tissue structures depicted therein that have the highest predictive value for membership of a particular class of image. This can generate new biomedical knowledge that has not been previously described in the biomedical literature. For example, by outputting and highlighting one of 1000 image tiles that show a particular tissue structure that is highly predictive of the image class label "cancer patient tissue that will benefit from treatment X," a correlation or causal relationship between the tissue structure depicted in this tile and a particular image class that may not have been previously discovered and published can be revealed. Furthermore, even if a tissue structure that predicts a particular image class is so known, a heat map or any other form of graphically representing the tiles with the highest predictive value can have the advantage of enabling a pathologist to more quickly and more accurately identify one or more tiles that contain the most interesting tissue structures with respect to a particular biomedical problem.

[0080] For example, image regions and each tile having a weighted predicted value wh that is very similar to the weighted predicted value of the highest scoring tile of the image can be represented in the relevance heat map with a first color (e.g., “red”) or high intensity value, and image regions and each tile whose weighted predicted value is not similar to the highest wh value of all tiles in the image can be represented in the relevance heat map with a second color different from the first color (e.g., “blue”) or low intensity value.

[0081] This can be advantageous because the GUI automatically calculates and presents a relevance heatmap showing the location and coverage of tissue regions and each image tile with high predictive power (or "prognostic value") wh weighted by the model's certainty value c for this tile. The relevance heatmap can highlight tissue regions with high relevance indicators. A tile is usually only a small sub-region of a whole slide image, and a heatmap and / or gallery of tiles sorted according to the tile's wh value may not provide an overview of the entire tissue sample. Overview information about the location and extent of tissue patterns with high predictive relevance can be provided by the relevance heatmap, which is advantageously combined with the original image of the whole slide tissue image in a highly intuitive and smart manner.

[0082] Computing and displaying a weighted predictive value-based heatmap can be advantageous because it indicates the predictive power of tiles for the endpoints used to train the MIL. Displaying the relevance heatmap to the user thus allows the user to quickly identify the location and extent of tiles that have tissue patterns that are predictive for a particular label within a full slide image.

[0083] According to an embodiment, the method further comprises a step of displaying the received images in a GUI on a screen, the images being grouped in the GUI into at least two different classes according to the results of the classification.

[0084] According to an embodiment, each of the at least two classes is selected from the group comprising: - Patients with responsiveness to certain drugs; - Patients who have developed metastases or certain forms of metastases (e.g., micrometastases); - Cancer patients who demonstrate a specific response to a particular treatment, e.g., pathological complete response (pCR); - Cancer patient tissues exhibiting specific morphological or microsatellite status; - Patients who have developed adverse reactions to certain medications; - Patients with a particular genetic attribute, e.g., a particular genetic signature; and / or - Patients with specific RNA expression profiles.

[0085] These labels can be useful in diagnosis, as well as in finding suitable drugs for treating a disease. However, the above labels are merely examples. Other patient-related attributes can also be used as labels (i.e., endpoints for training the MIL program), as described above. The term "patient-related" can also include the term treatment-related, since the effectiveness of a particular treatment for a disease is also related to the patient being treated.

[0086] Embodiments of the present invention are used to identify tissue patterns indicative of patient-related attribute values. This method may be advantageous because it may combine the advantages of image analysis methods based on the explicit knowledge of biomedical experts with the advantages of machine learning methods.

[0087] According to an embodiment, the certainty value based pooling function is a certainty value based max pooling, average pooling or attention pooling function.

[0088] According to an embodiment, the method further comprises: providing a trained attention MIL that has learned which tile-derived feature vectors are most relevant for predicting a tile's class membership; - calculating an attention weight (aw) for each of the tiles as a function of the feature vector of each tile by the attention MIL, the attention weight being a measure of the relevance of the feature values ​​of this tile with respect to the membership of this tile to a class; multiplying the attention weight (aw) of the tile by the feature vector value of the tile to obtain an attention-based feature vector having an attention-weighted feature value for the tile; using the attention-based feature vector as a feature vector that is input to a MIL program to calculate a prediction value (h), a certainty value (c) and / or a weighted prediction value (wh) for the slide, whereby the prediction value (h), the certainty value (c) and / or the weighted prediction value (wh) are calculated as the attention-based prediction value, the attention-based certainty value and / or the attention-based weighted prediction value; or - further including multiplying the attention weight (aw) for the tile by the prediction value (h), the certainty value (c) for the tile, or the weighted prediction value (wh) for the tile calculated by the MIL to obtain an attention-based prediction value, an attention-based certainty value, and / or an attention-based weighted prediction value.

[0089] For example, the attention weights can be calculated as described in MAXIMILIAN ILSE et al.: “Attention-based Deep Multiple Instance Learning”, ARXIV.ORG, CORNELL UNIVERSITY LIBRARY, 201 OLIN LIBRARY CORNELL UNIVERSITY ITHACA, NY 14853, February 13, 2018, XP081235680, which describes the application of an attention-based multi-instance learner (MIL) to histopathology datasets, e.g., as described with reference to Equation (9).

[0090] The combination of attention weights and certainty values ​​can be particularly advantageous when the image data is noisy and / or significantly different from the images used during the training phase, since the calculation of attention weights alone may provide insufficient (inaccurate) classification results. However, the drawbacks and lack of robustness of attention-based weights can be compensated for by also taking into account certainty values, so that even if a high attention weight is obtained for a particular vector / tile, if the certainty value obtained for this tile is low, the contribution of this tile / vector is largely ignored, thereby avoiding false positive classification results.

[0091] Please note that "attention weight" and "certainty value" relate to completely different concepts: "certainty value" indicates the certainty value of a predictive model regarding the contribution of a vector / tile to the classification result, while "attention weight" indicates the relevance (predictive power) of a feature vector / tile to the classification result. For example, a first feature vector may be derived from a first tile of an image, and a second vector may be derived from a second tile of the same image.

[0092] For example, the attention MIL may determine that the image shows "tumor tissue" based on the first feature vector because the first feature vector indicates that the first tile contains a first pattern that is highly correlated / has a high predictive value with the tissue type class "tumor tissue." Thus, the attention MIL calculates a high attention weight for the first tile because the identified pattern is highly predictive of the class "tumor tissue."

[0093] In addition, because the second feature vector indicates that the second tile contains another pattern that is weakly positively correlated with the classification "tumor tissue," the attention MIL can determine that the image is indicative of "tumor tissue" based on the second feature vector. This means that about 60% of all images that contain tiles with this pattern are actually tumor tissue images, and about 40% of those images may contain tiles with this pattern but still be indicative of "healthy" tissue.

[0094] If certainty is not calculated / considered, the following situation may occur: the correlation of the first pattern with the "tumor class label" may be strong (all training images with at least one tile with the first pattern are indeed tumor tissue images, and there are no "healthy tissue" training images that contain a tile with this first pattern), but the certainty value of the classification "tumor tissue" may be low. For example, the total number of tiles with the first pattern in the training set may be low, and therefore the "certainty" of any prediction based on this first pattern may be low. Thus, the attention MIL may calculate a high predictive relevance of a tile with respect to a particular class label, but the certainty value calculated for this tile is low, and vice versa. For the second tile, the attention weighting will be low since the second pattern appears to be a poor predictor of class membership, and the certainty value may be high or low depending on the quality and abundance of examples in the training dataset that contain the second pattern.

[0095] Taking into account both attention weights and certainty values ​​ensures that patterns in the training dataset that appear to be strongly correlated / highly predictive of a particular class membership are ignored, or at least downweighted, if the training data basis for this prediction, represented by the certainty value of the model computed for this prediction / classification, is low.

[0096] For example, a max pooling function may classify an image as being a member of a particular class only if the attention-based prediction value weighted by the certainty value exceeds a given threshold.

[0097] According to an embodiment, the predicted value h and / or the weighted predicted value wh calculated for each tile can be output together with the graphic representation of the associated tile in the gallery. For example, the tiles in the gallery can be sorted according to the numerical value wh calculated for each tile. In this case, the location of the tiles in the gallery allows a pathologist or other user to identify tissue patterns depicted in some of the tiles that have been found to be highly predictive for a particular label. Additionally or alternatively, the numerical values ​​can be displayed in spatial proximity to their respective tiles, allowing the user to examine and understand the tissue patterns of tissues depicted in one or more tiles with similar numerical values ​​for a particular label.

[0098] The gallery can be output based on training images at the end of the training phase and / or based on test images at testing. The image tile gallery generated as the output of the trained MIL program can reveal predictive tissue signatures for certain patient-related attribute values ​​of the patient. Presenting numerical values ​​in combination with the image tiles can have the advantage that, at least in many cases, predictive tissue patterns (which may also be referred to as "tissue signatures") can be identified and verbalized by a pathologist by comparing some tiles in the gallery that have similar numerical values ​​with other tiles that have higher or lower numerical values, and by comparing the tissue signatures depicted in a subset of those tiles in the report gallery.

[0099] In a further aspect, the present invention relates to an image analysis system for classifying tissue images, the image analysis system comprising at least one processor and a volatile or non-volatile storage medium containing digital images each representative of a tissue sample from a patient.

[0100] The image analysis system further comprises an image segmentation module executable by the at least one processor and configured to segment each of the images into a set of image tiles.

[0101] The image analysis system further comprises a feature extraction module executable by the at least one processor and configured to calculate, for each of the tiles, a feature vector including image features selectively extracted from said tiles.

[0102] The image analysis system further includes a multi-instance learning (MIL) program executable by at least one processor and configured to use a model to classify any input image as a member of one of at least two distinct classes based on feature vectors extracted from all tiles of said input image. The MIL program comprises: - calculating, for each of the tiles, a certainty value, the certainty value indicating the certainty of the model regarding the contribution of the feature vector of the tile to the classification of the image from which the tile was derived; -For each image, using a certainty value based pooling function for aggregating feature vectors extracted from the image into a global feature vector as a function of the certainty values ​​of tiles of the image, by the MIL program, and calculating an aggregate prediction value (herein called "ah" according to an embodiment of the present invention) from the global feature vector; or calculating a prediction from each of the image's feature vectors using a MIL program, and using a certainty-based pooling function to aggregate the image's predictions into an aggregate prediction ("ah") using the MIL program as a function of the certainty values ​​of the tiles of the image; - classifying each of the images as a member of one of at least two different classes based on the aggregate prediction value.

[0103] According to an embodiment, the system comprises an interface for outputting the classification results. For example, the interface can be a machine-to-machine interface, e.g., an application program interface for sending the classification results to a software application program. Additionally or alternatively, the interface can be or include a man-machine interface, e.g., a screen configured to display a GUI including a graphical representation of the classification results. For example, the MIL program can be configured to output the classification results to a user via the GUI.

[0104] According to an embodiment, the GUI allows the user to select whether the heatmap, sometimes referred to as a "relevance heatmap," is computed based on the tile's predicted value h or based on the tile's weighted predicted value wh, allowing the user to evaluate the effect of taking into account model uncertainty.

[0105] Feature Extraction Approach According to an embodiment, the calculation of a feature vector for each of the tiles at training and / or testing time comprises extracting one or more image features from the tile and representing the extracted features in the form of one or more features in the feature vector. Optionally, the calculation of the feature vector further comprises receiving patient-related data of a patient whose tissue sample is depicted in the tile and representing the patient-related data in the form of one or more features in the feature vector. The patient-related data can be, for example, genomic data, RNA sequencing data, the patient's known diseases, age, sex, metabolite concentrations in body fluids, health parameters and current medications.

[0106] According to an embodiment, computing the feature vector is performed by a trained machine learning logic, in particular a trained fully convolutional neural network including at least one bottleneck layer.

[0107] According to an embodiment, the trained machine learning logic used for feature extraction ("feature extraction MLL") is trained in a supervised manner employing a MLL of the type of fully convolutional network, including a bottleneck, such as UNET. The "Unet" architecture is described by Olaf Ronneberger, Philipp Fischer, and Thomas Brox in "U-Net: Convolutional Networks for Biomedical Image Segmentation", Computer Science Department and BIOSS Centre for Biological Signalling Studies, University of Freiburg, Germany (arXiv:1505.04597v1 18 May 2015). This document is available for download from the Cornell University Library at https: / / arxiv.org / abs / 1505.04597.

[0108] For example, the feature extraction MLL can be trained to perform a tissue image segmentation task whereby the segments identified include two or more of the following tissue image segment types: tumor tissue, healthy tissue, necrotic tissue, tissue containing specific objects such as tumor cells, blood vessels, stroma, lymphocytes, etc., and background areas. According to some embodiments, the feature extraction MLL is trained in a supervised manner by using a classification network such as Resnet, ImageNet, or SegNet to train it to classify tiles of images having specific pre-defined classes or objects.

[0109] After training the feature extraction MLL, the MLL is split into an "encoder" section (including an input layer, one or more intermediate layers, and a bottleneck layer) and a "decoder", i.e., output generating section. The "encoder" section of the trained MLL up to the bottleneck layer is used to extract and compute feature vectors for each input tile according to an embodiment of the present invention. A bottleneck layer is a layer of a neural network that contains significantly fewer neurons than the input layer. For example, a bottleneck layer can be a layer that contains less than 60% or even less than 20% of the "neurons" of the input layer. The number and ratio of neurons in different layers can vary greatly depending on different network architectures. A bottleneck layer is a hidden layer.

[0110] According to one example, the network of the feature extraction MLL has a network architecture based on UNET, which has an input layer with 512*512*3 (512×512RGB) neurons and a bottleneck layer with 9*9*128 neurons. Therefore, the number of neurons in the bottleneck layer is about 1.5% of the number of neurons in the input layer.

[0111] According to one example, the network of the feature extraction MLL has a Resnet architecture that implements a supervised or unsupervised learning algorithm: the input layer contains 512x512x3 neurons, and the bottleneck layer and the corresponding feature vector output by the bottleneck layer typically contain 1024 or 2048 elements (neurons / number).

[0112] According to an embodiment, feature extraction is performed by a feature extractor module based on the ResNet-50 (He et al., 2016) architecture trained on the ImageNet natural image dataset. Some detailed examples of feature extraction from images based on this architecture are described in Pierre Courtiol, Eric W. Tramel, Marc Sanselme, and Gilles Wainrib: "CLASSIFICATION AND DISEASE LOCALIZATION IN HISTOPATHOLOGY USING ONLY GLOBAL LABELS: A WEAKLY-" SUPERVISED APPROACH", arXiv:1802.02212, https: / / arxiv.org / pdf / 1802.02212.pdf, submitted February 1, 2018, available online via the Cornell University Library.

[0113] According to an embodiment, the output generated by one of the layers of the trained feature extraction MLL for a particular tile is used as the feature vector extracted from the tile by the MIL program. This one layer can be, in particular, a bottleneck layer. According to an embodiment, the feature extraction MLL is trained in an unsupervised or self-supervised manner, as described in Mathilde Caron and Piotr Bojanowski and Armand Joulin and Matthijs Doze: "Deep Clustering for Unsupervised Learning of Visual Features", CoRR, 1807.05520, 2018, available electronically via https: / / arxiv.org / abs / 1807.05520.

[0114] Alternatively, the feature extraction MLL can be trained according to "Unsupervised Representation Learning by Predicting Image Rotations" by Spyros Gidaris, Praveer Singh, and Nikos Komodakis, ICLR 2018 Conference, February 15, 2018, available electronically via https: / / openreview.net / forum?id=S1v4N2l0-.

[0115] Further alternatively, the feature extraction MLL can be trained according to Elad Hoffer, Nir Ailon, "Semi-supervised deep learning by metric embedding," November 4, 2016, ICLR 2017, available electronically via https: / / openreview.net / forum?id=r1R5Z19le.

[0116] The dataset for training the feature extraction MLL can be another tissue image dataset and / or a set of tissue images that are subsequently used to train the MIL program. Because the feature extraction MLL is trained to identify tissue types and respective image segments, rather than patient-related attribute values ​​of patients that are used as endpoints of the learning phase of the MIL program, any labels associated with the training images are not evaluated or otherwise used by the feature extraction MLL in the training phase.

[0117] A feature extraction approach that allows the use of proximity-based similarity labels According to an embodiment, a feature vector is computed by a feature extraction machine learning logic ("feature extraction MLL") trained on a training dataset containing labeled tile pairs, whereby each label represents the similarity of the two tissue patterns depicted by the tile pair and is computed as a function of the spatial distance of the two tiles in the tile pair.

[0118] According to a preferred embodiment, labels are assigned to tile pairs in the training dataset fully automatically.

[0119] This approach can be beneficial for several reasons: the spatial proximity of two image regions is a feature that is always and inherently available in all digital images of tissue samples. The problem is that the spatial proximity of the images and the respective tissue regions themselves typically does not reveal any relevant information regarding biomedical problems such as tissue type classification, disease classification, prediction of the persistence of a particular disease, or image segmentation tasks. Applicant has surprisingly observed that the information provided in the spatial proximity of two image regions ("tiles") is an accurate indicator of the similarity of the two image regions, at least when a large number of tiles and their respective distances are analyzed during the training phase of the MLL. Thus, by making available the inherently available information "spatial proximity" of the two tiles, a large annotated dataset can be automatically provided that can be used to train the MLL to automatically assign tissue pattern similarity labels to the two compared tiles. The trained MLL can be used to automatically determine whether two images or image tiles received as input depict similar or dissimilar tissue patterns. However, the dataset can be used for other more complex tasks such as image similarity search, image segmentation, tissue type detection, and tissue pattern clustering. Thus, the applicant surprisingly observed that the information provided in the spatial proximity of tiles can be used to automatically generate annotated training data that can enable training of an MLL to reliably determine image similarity and, in addition, output feature vectors that can be used by additional data processing units for multiple complex image analysis tasks in digital pathology. None of these approaches require domain experts to manually annotate the training data.

[0120] When a training image containing many different tissue patterns (e.g., "non-tumor" and "tumor") is divided into many different tiles, the shorter the distance between two tiles, the higher the probability that both of the compared tiles depict the same tissue pattern, such as "non-tumor". However, there are some tile pairs near the boundary of two different patterns depicting different tissue patterns (e.g., the first tile is "tumor" and the other is "non-tumor"). These tile pairs generate noise because they depict different tissue patterns, even though they are in close spatial proximity. Applicant surprisingly observed that this noise generated by tile pairs spanning the boundary between different tissue patterns, combined with the simplifying assumption that spatial proximity indicates the similarity of the depicted tissue patterns, does not significantly degrade the accuracy of the trained MLL. In fact, Applicant observed that the accuracy of the MLL trained according to an embodiment of the present invention can surpass those of existing benchmark methods.

[0121] In a further beneficial aspect, training data can now be rapidly, completely, and automatically generated for many different sets of images. Currently, there is a lack of available annotated datasets that capture the natural and realistic variability in histopathology images. Even existing large datasets, such as Camelyon, consist of only one type of stain (hematoxylin and eosin) and one type of cancer (breast cancer). The texture and object shape of histopathology images can vary greatly in images from different cancer types, different tissue stain types, and different tissue types. Additionally, histopathology images contain many different textures and object types with different domain-specific meanings (e.g., stroma, tumor-infiltrating lymphocytes, blood vessels, fat, healthy tissue, necrosis, etc.). Thus, embodiments of the present invention can enable the automatic generation of annotated datasets for each of multiple different cancer types, cancer subtypes, staining methods, and patient groups (e.g., treated / untreated, male / female, older / younger than a threshold age, biomarker-positive / biomarker-negative, etc.). Thus, embodiments of the present invention can automatically generate annotated training data and train each MLL on the training data, such that the resulting trained MLL is adapted to precisely address the biomedical problem in a specific manner for each of a plurality of different patient groups. Contrary to current approaches, where an MLL trained on a manually annotated breast cancer dataset provides suboptimal results for colon cancer patients, embodiments of the present invention can enable MLLs to be generated individually for each of the different patient groups.

[0122] According to an embodiment, the label indicating the degree of similarity of the two texture patterns is a binary data value, i.e., a value that can have one of two possible options. For example, the label can be "1" or "similar", indicating that the two tiles depict similar texture patterns. Alternatively, the label can be "0" or "dissimilar", indicating that the two tiles depict dissimilar texture patterns. According to other embodiments, the label can be more fine-grained, and can be a data value selected from a limited set of three or more data values, such as "dissimilar", "similar", and "highly similar". According to yet other embodiments, the label can be more fine-grained, and can be a numerical value whose magnitude positively correlates with the degree of similarity. For example, the numerical value can be calculated as a function that linearly converts the spatial distance between the two tiles in the pair to a numerical value representing the texture pattern similarity and vice versa. The greater the spatial distance, the smaller the numerical value indicating the texture pattern similarity. A wide variety of MLL architectures exist, which can process and use different types of labels in the training dataset (e.g., ordinal or numeric). The type of MLL is chosen such that it can process the automatically generated labels of the training dataset.

[0123] According to an embodiment, the MLL used for feature extraction, trained on an automatically annotated training data set, is adapted to learn according to a supervised learning algorithm. Supervised learning concerns finding a mapping that transforms a set of input features as members of one or more output data values. The output data values ​​are provided during training as labels, for example as binary option labels "similar" or "dissimilar", or as numerical values ​​that are quantitative measures for similarity. In other words, during training, the data values ​​that have to be predicted are explicitly provided to the model of the MLL in the form of labels of the training data. Supervised learning suffers from the problem that the training data needs to be labeled in order to define the output space for each sample.

[0124] According to an embodiment, at least some or all of the tile pairs each depict two tissue regions contained in the same tissue slice, each of the tissue slices depicted in a respective one of the received digital images, and the distance between the tiles is calculated in a 2D coordinate system defined by the x and y dimensions of the received digital image from which the tiles in the pair are derived.

[0125] According to an embodiment, the tile pairs are generated by randomly selecting tile pairs within each of a plurality of different images. The random-based selection ensures that the spatial distance between the tiles in each pair is different. A similarity label, e.g., in the form of a numerical value that is inversely correlated with the distance between the two tiles, is calculated and assigned to each pair.

[0126] According to another embodiment, the tile pairs are generated by selecting at least some or all of the tiles of each received image as a starting tile, selecting for each starting tile all or a predetermined number of "nearby tiles", the "nearby tiles" being tiles within a first circle centered on the starting tile, whereby the radius of this circle is equal to a first spatial proximity threshold, and selecting for each starting tile all or a predetermined number of "far tiles", the "far tiles" being tiles outside a second circle centered on the starting tile, whereby the radius of said circle is equal to a second spatial proximity threshold, the selection of the predetermined number being made by randomly choosing this number of tiles within the respective image area. The first and second proximity thresholds may be the same, but it is preferred that the second proximity threshold is greater than the first proximity threshold. For example, the first proximity threshold may be 1 mm and the second proximity threshold may be 10 mm. A first set of tile pairs is then selected, each tile pair including a starting tile and a nearby tile located within the first circle. Each tile pair in the first set is assigned an organizational pattern with the label "similar." In addition, a second set of tile pairs is selected, such that each pair in the set includes the starting tile and one of the "distant tiles." Each tile pair in the second set is assigned an organizational pattern with the label "dissimilar." For example, this embodiment can be used to generate "binary" labels of "similar" or "dissimilar."

[0127] According to embodiments, the distance between tiles is measured in a 2D coordinate system defined by the x-axis and y-axis of the digital image from which the tiles are derived. These embodiments can be used in situations where multiple tissue sample images depicting different patient tissue samples and / or different regions within the same patient are available, whereby the different regions are separated from each other or whereby the exact positions of the two regions with respect to each other are unknown. In this case, the spatial proximity between tiles is measured only in the 2D plane of pixels defined by the digital images. Based on the known resolution factor of the image acquisition device (e.g., a microscope camera or a slide scanner), the distance between tiles of the original image can be used to calculate the distance between the tissue regions in the tissue sample depicted by the two tiles.

[0128] According to an embodiment, at least some or all of the tile pairs depict two tissue regions contained in two different tissue slices of a bundle of adjacent tissue slices, each of the tissue slices being depicted in a respective one of the received digital images. The received images depicting the tissue slices of the bundle of adjacent tissue slices are aligned with each other in a 3D coordinate system. The distance between the tiles is calculated in the 3D coordinate system.

[0129] For example, some or all of the received digital images may depict tissue samples that are slices within a tissue block of adjacent tissue slices. In this case, the digital images may be aligned with one another in a common 3D coordinate system such that the positions of the digital images in the 3D coordinate system reproduce the positions of the tissue slices that each depicted within the tissue block. This may allow for determining the distance of the tiles in the 3D coordinate system. The selection of "nearby" and "far" tiles may be performed as described above for the 2D coordinate system case, with the only difference being that the tiles in at least some tile pairs are derived from different ones of the received images.

[0130] According to some embodiments, the annotated training data includes both tile pairs derived from the same digital image, as well as tile pairs derived from different images aligned with each other in a common 3D coordinate system. This can be beneficial because consideration of the third dimension (spatial proximity of tiles representing tissue regions in different tissue samples) can greatly increase the number of tiles in the training data when only a few images of each tissue sample are available, which may be beneficial because the tissue samples belong to the same cell block, such as a 3D biopsy cell block.

[0131] According to an embodiment, each tile depicts a tissue or background region having a longest edge length of less than 0.5 mm, preferably less than 0.3 mm.

[0132] A small tile size can have the advantage of reducing the number and percentage of area of ​​tiles that depict a mix of different texture patterns. This can help reduce noise generated by tiles that depict two or more different texture patterns, and by tile pairs near "texture pattern boundaries" that depict two different texture patterns. In addition, a small tile size can make it possible to generate and label a large number of tile pairs, thereby increasing the amount of labeled training data.

[0133] According to an embodiment, the automatic generation of tile pairs includes: generating a first set of tile pairs using a first spatial proximity threshold, where the two tissue regions depicted by the two tiles of each tile pair in the first set are separated from each other by a distance smaller than the first spatial proximity threshold; and generating a second set of tile pairs using a second spatial proximity threshold, where the two tissue regions depicted by the two tiles of each tile pair in the second set are separated from each other by a distance larger than the second spatial proximity threshold. For example, this can be performed by selecting a number of starting tiles, calculating first and second circles around each starting tile based on the first and second spatial proximity thresholds, and selecting tile pairs including the starting tile and a "nearby tile" (first set) or a "far tile" (second set), as already described above for an embodiment of the present invention.

[0134] According to an embodiment, the first and second spatial proximity thresholds are the same, for example 1 mm.

[0135] According to a preferred embodiment, the second spatial proximity threshold is at least 2 mm greater than the first spatial proximity threshold, which may be preferred in cases where the texture patterns change gradually from one pattern to another, as the difference between the texture patterns depicted in the "distant tiles" compared to the texture patterns depicted in the "nearby" tiles may be clearer and the learning effect may be improved.

[0136] According to an embodiment, the first spatial proximity threshold is a distance less than 2 mm, preferably less than 1.5 mm, in particular 1.0 mm.

[0137] Additionally or alternatively, the second spatial proximity threshold is a distance of greater than 4 mm, preferably greater than 8 mm, and in particular 10.0 mm.

[0138] These distance thresholds refer to the distance of the digital image and the tissue region (or slice background region) depicted in each tile. Based on the known magnification of the image acquisition device and the resolution of the digital image, this distance can be converted to a distance in the 2D or 3D coordinate system of the digital image.

[0139] For example, the distance between tiles (and the tissue regions depicted therein) can be measured, for example, between the centers of two tiles in a 2d or 3D coordinate system. According to an alternative implementation variant, the distance is measured between the two tile edges (image region edges) that are closest to each other in the 2D or 3D coordinate system.

[0140] The above thresholds have been observed to provide labeled training data that allows for automatically generating a trained MLL that can accurately identify similar and dissimilar tissue patterns for breast cancer patients. In some other embodiments, the first and second spatial proximity thresholds can have other values. In particular, in cases where different sets of received digital images showing different tissue types or cancer types are used, the first and second spatial proximity thresholds can have other values ​​different from the distance thresholds provided above.

[0141] According to an embodiment, the method further includes generating a training data set for training the feature extraction MLL, the method including: receiving a plurality of digital training images, each depicting a tissue sample; dividing each of the received training images into a plurality of tiles ("feature extraction training tiles"); automatically generating tile pairs, each tile pair being assigned a label indicating the degree of similarity of the two tissue patterns depicted in the two tiles of the pair, the degree of similarity being calculated as a function of the spatial proximity of the two tiles in the pair, with distance positively correlated with dissimilarity; training a machine learning logic - the MLL - using the labeled tile pairs as training data to generate a trained MLL, the trained MLL having been learned to extract feature vectors from digital tissue images representing images such that similar images have similar feature vectors and dissimilar images have dissimilar feature vectors; and using the trained MLL, or a component thereof, as the feature extraction MLL used to compute the feature vectors of the tiles.

[0142] This approach can be beneficial because the labels of the training dataset can be automatically generated based on the information inherently contained in all digital pathology images. Thus, by simply selecting training images, an annotated dataset can be generated for training a feature extraction MLL that is specifically adapted to the biomedical problem currently being addressed. All further steps, such as segmentation, labeling and machine learning steps, can be performed fully automatically or semi-automatically.

[0143] According to an embodiment, the trained MLL is a Siamese network that includes two neural sub-networks that are connected by their output layers. One of the sub-networks of the trained Siamese network is stored separately on a storage medium and is used as a component of the trained MLL that is used to calculate the feature vector of a tile.

[0144] According to embodiments, the MIL program learns in the training phase to convert feature vectors into prediction values ​​h that can represent the probability of a particular bag label (i.e., membership in an image class). The labels can represent classes (e.g., patients responding to a treatment with a particular drug D or a numerical range indicating the degree of response). This learning can be described mathematically as learning a nonlinear transformation function that converts feature values ​​into one of the labels provided during training. According to some embodiments, at test time, some minor structural changes (e.g., disabling the dropout layer) are applied to the trained MIL program, and no sampling of the test data is performed. The main change when applying the trained MIL program at test time is that all instances (tiles) in the bag of test data are analyzed by the MIL program and a final numerical value indicating the predictive power is calculated for each of the tiles and for each of the multiple labels provided in the training phase. Finally, a final numerical value is calculated for the entire image or for a particular patient by aggregating the (weighted) prediction values ​​calculated for the tile of the image for the multiple labels. The end result of applying the trained MIL program to one or more images of a patient is one of the labels with the highest probability (e.g., "Patient responds to treatment with drug D!"). In addition, one of the tiles with the highest predictive power for this label can be presented in a report image tile gallery that is structurally equivalent to the report image tile gallery described above for the training phase.

[0145] According to an embodiment, the method further includes automatically selecting or allowing a user to select one or more "high predictive power tiles", which are tiles whose predicted or weighted predicted values, indicative of the predictive power of that feature vector for a particular one of the labels, exceed a high predictive power threshold.

[0146] Additionally or alternatively, the method further includes automatically selecting or allowing a user to select one or more "artifact tiles" that depict one or more artifacts or whose numerical value indicating the predictive power of the feature vector for a particular one of the labels is below a minimum predictive power threshold.

[0147] In response to the selection of one or more high predictive power tiles and / or artifact tiles, the MIL program is automatically retrained, thereby excluding the high predictive power tiles and artifact tiles from the training set.

[0148] These features can have the advantage that the retrained MIL program can be more accurate, since the excluded artifact tiles are no longer considered during retraining. Thus, any bias in the learned transformation caused by tiles in the training dataset that depict artifacts is avoided and removed by retraining the MIL program on a reduced version of the training dataset that does not include the artifact tiles.

[0149] Allowing users to prune highly predictive tiles from the training dataset may be counter-intuitive, but nevertheless provides an important benefit: sometimes, for some labels, the predictive power of some tissue patterns is trivial.

[0150] For example, a tissue section containing many tumor cells expressing lung cancer specific biomarkers is, of course, an important predictive marker for the presence of lung cancer disease. However, a pathologist may be more interested in some less obvious tissue patterns, such as the presence and / or location of non-tumor cells, e.g., FAP+ cells.

[0151] According to an embodiment, the prediction value h calculated by the MIL program for a particular tile based on the feature vector of this tile indicates the predictive power of the tile for the label of the bag (image).

[0152] According to some embodiments, the weighted prediction value wh of a tile is calculated by multiplying the prediction value h of this tile by the certainty value c of the model calculated for the tile.

[0153] According to another embodiment, the method comprises calculating, for each of the tiles, a feature vector in the form of a weighted feature vector calculated as a function of the weight calculated by the feature extraction program as the certainty value c of the model of said tile and of the feature vector calculated for said tile. In particular, the certainty value c can be multiplied with the feature vector of this tile.

[0154] According to one embodiment, the MIL program is trained such that the model certainty values ​​cs calculated for the tiles of the bag are provided as input to the MIL program along with the feature vectors of the tiles. The MIL is trained such that the MIL learns more from tiles whose feature vectors have higher certainty values ​​(i.e., "weights") than from tiles whose feature vectors have lower weights. In other words, during training, the MIL program learns to correlate the influence of tiles and their feature vectors on the predicted value of the tile with the certainty values ​​calculated for the particular tile, which are used as weights.

[0155] "Tissue sample" as used herein is a 2D or 3D assembly of cells that can be analyzed by the method of the present invention. The cell aggregate can be a slice of an ex vivo cell block. For example, the sample can be prepared from tissue collected from a patient, for example, liver, lung, kidney, or colon tissue samples from a cancer patient. The sample can be a whole tissue or a TMA section on a microscope slide. Methods for preparing tissue samples placed on slides are well known to those skilled in the art and suitable for use in the present invention.

[0156] The tissue sample can be stained using any reagent or biomarker label, such as a dye or staining agent, or a histochemical or immunohistochemical that reacts directly with a specific biomarker or with various types of cells or subcellular compartments. Not all stains / reagents are compatible. Thus, the type of stain employed and its sequence of application should be well considered, which can be easily determined by a person skilled in the art. Such histochemicals may be chromophores detectable by transmission microscopy or fluorophores detectable by fluorescence microscopy. In general, a sample containing cells can be incubated with a solution containing at least one histochemical that reacts directly with or binds to a chemical group of interest. Some histochemicals are typically incubated together with a color-blocking agent or metal that allows staining. A sample containing cells can be incubated with a mixture of at least one histochemical that stains a component of interest and another histochemical that acts as a counterstain and binds to an area outside the component of interest. Alternatively, a mixture of multiple probes can be used in the staining, providing a method of identifying the location of a particular probe. Procedures for staining samples containing cells are well known to those of skill in the art.

[0157] As used herein, an "image analysis system" is a system, e.g., a computer system, adapted to evaluate and process digital images, particularly images of tissue samples, to assist a user in evaluating or interpreting, e.g., classifying, the images and / or to extract biomedical information implicitly or explicitly contained within the images. For example, the computer system can be a standard desktop computer system or a distributed computer system, such as a cloud system. In general, a computerized histopathology image analysis system takes as its input a single or multi-channel image captured by a camera and attempts to provide additional quantitative information for assistance in diagnosis or treatment. The image can be received directly from the camera or can be read from a local or remote storage medium.

[0158] The embodiments of the present invention can be used to classify tissue images, for example, to perform image-based tumor staging and / or to determine which patient subgroups of a larger patient population are likely to benefit from a particular drug. Personalized medicine (PM) is an emerging medical field that aims to provide effective, targeted treatment strategies based on an individual's genomic, epigenomic, and proteomic profiles. PM not only attempts to treat patients, but also protects them from negative side effects of less effective treatments. Some mutations that often occur with tumor progression cause resistance to certain treatments. Thus, the patient's variability profile, which may be revealed at least in part by tissue images of biomarker-specific stained tissue samples, allows a trained MIL program to clearly determine whether a particular treatment will be effective for an individual patient. Currently, a trial-and-error approach is required to determine whether a prescribed drug will be effective for a patient. This trial and error process can have many negative side effects, such as undesirable complex drug interactions, frequent changes in prescribed drugs, long delays in identifying effective drugs, and disease progression. Image classification performed according to embodiments of the present invention can be used to stratify individuals into subpopulations that differ in their response to therapeutic agents for their particular disease. For example, several ALK kinase inhibitors are useful drugs to treat about 5% of NSCLC lung cancer patients with increased expression in the ALK gene. However, after some time, the kinase inhibitors become ineffective due to mutations in the ALK gene or other genes downstream of the ALK signaling cascade. Thus, intelligent molecular characterization of lung cancer patients allows optimal use of several mutation-specific drugs through patient stratification. Thus, the "patient group" from which training or test images are obtained can be groups such as "100 breast cancer patients", "100 HER+ breast cancer patients", "200 colon cancer patients", etc.

[0159] A "digital image" as used herein is a numerical representation, usually binary, of a two-dimensional image. Typically, tissue images are raster type images, meaning that the image is a raster ("matrix") of pixels, each of which is assigned at least one intensity value. Some multi-channel images may have pixels with one intensity value per color channel. A digital image contains a fixed number of rows and columns of pixels. A pixel is the smallest individual element in an image and holds a traditional value that represents the brightness of a given color at any particular point. Typically, pixels are stored in computer memory as a raster image, or raster map, which is a two-dimensional array of small integers. These values ​​are often transmitted or stored in compressed form. Digital images can be acquired by digital cameras, scanners, coordinate measuring machines, microscopes, slide scanning devices, and the like.

[0160] As used herein, a "label" is a data value, e.g., a string or a number, that represents and designates a patient-related attribute value and / or a class of patients having this attribute value. Examples of labels can be "Patient Response to Drug D = True", "Patient Response to Drug D = False", "Progression Free Survival > 6 months", "Patient has micrometastases", etc.

[0161] An "image tile" or "tile" as used herein is a sub-region of a digital image. In general, tiles generated from a digital image can have any shape, such as circular, elliptical, polygonal, rectangular, square, etc., and can be overlapping or non-overlapping. According to a preferred embodiment, tiles generated from an image are rectangular, and preferably overlapping tiles. Using overlapping tiles can also have the advantage that tissue patterns that would otherwise be fragmented by the tile generation process are represented in the bag. For example, the overlap of two overlapping tiles can be 20 to 30%, e.g., 25%, of the area of ​​a single tile.

[0162] A "pooling function" is a permutation invariant transformation that generates a single aggregated numerical value for a bag of tiles based on all tiles. The pooling function may allow to specify how the information encoded in all tiles of a bag is taken into account during the training and / or testing phases for computing an image classification result. A pooling function is used by the MIL program in the training phase, and the same or a different pooling function is used by the trained MIL program in the testing phase. According to a preferred embodiment, the pooling function used by the MIL program at training and testing time is a certainty value-based pooling function. For example, the pooling function used at training time is a certainty value-based average pooling function, and the pooling function used at testing time is a certainty value-based max pooling function.

[0163] A "certainty value based pooling function" is a pooling function that explicitly or implicitly takes into account the certainty values ​​calculated by the MIL program for each of the tiles of the image.

[0164] A "feature vector" as used herein is a data structure that contains information that describes the characteristics of an object. The data structure can be a single or multi-dimensional data structure, where a particular type of data value is stored at each of the locations in the data structure. For example, the data structure can be a vector, an array, a matrix, etc. A feature vector can be considered as an n-dimensional vector of numerical features that represent some object. In image analysis, features can take many forms. A simple feature representation of an image is the raw intensity value of each pixel. However, more complex feature representations are possible. For example, features extracted from an image or image tile can also be SIFT descriptor features (scale invariant feature transforms). These features capture the spread into different line orientations. Other features can indicate contrast, gradient orientation, color composition, and other aspects of the image or image tile.

[0165] An "embedding" is a transformation of a high-dimensional vector into a low-dimensional space. Ideally, an embedding captures some of the semantics of the input by placing semantically similar inputs close to each other in the embedding space. Within the context of this application, an embedding is considered to be a type of feature vector, since it is derived by a feature vector that is first extracted by a mathematical transformation, and therefore the parameter values ​​in the resulting embedding can also be thought of as "feature values" extracted from a particular tile.

[0166] As used herein, "machine learning logic (MLL)" is program logic, such as software, e.g., neuronal networks or support vector machines, that has been or can be trained in a training process and that includes a predictive model that has learned to perform some predictive and / or data processing task (e.g., image classification) based on the training data provided as a result of the training phase. Thus, MLL can be program code that is not explicitly specified, at least in part, by the programmer, but that has been implicitly learned and modified in a data-driven learning process that builds one or more implicit or explicit models from sample inputs. Machine learning can employ supervised or unsupervised learning.

[0167] The expression "multiple instance learning" (MIL) as used herein refers to a kind of (weakly) supervised machine learning approach. Instead of receiving a set of individually labeled instances, the learner receives a set of labeled bags, each containing many instances. In the simple case of multiple instance binary classification, a bag can be labeled negative if all instances in the bag are negative. On the other hand, if at least one instance in the bag is positive, the bag is labeled positive. From the collection of labeled bags, the learner tries to either (i) introduce a concept that correctly labels the individual instances, or (ii) learn how to label the bag without introducing the concept. A useful and simple example of MIL is given in Babenko, Boris, "Multiple instance learning: algorithms and applications" (2008). However, MIL programs according to some embodiments also cover training based on more than two different labels (endpoints).

[0168] According to an embodiment of the present invention, the MIL program is used to calculate a predictive value for each instance (tile) of a bag (multiple tiles or preferably all tiles of a digital tissue image), and therefore also for the tissue pattern depicted in the tile. In this step, new biomedical knowledge can be identified by the MIL program, because in the training data, the label of the image and the respective tile are given as the end point for training, and it is not the individual features of the feature vector derived from the tile that are strongly (positively or negatively) correlated with the label and therefore predictive for this label.

[0169] In other words, MIL refers to a form of machine learning (weakly supervised) in which training instances belonging to a common entity are provided with a set of instances, the set being called a bag. During training, the entire bag is provided with a label ("class label") indicating the membership of the bag in a particular class, but the labels of the individual instances are not known. For example, digital training images, whose membership in classes is known, can be divided into image tiles. The training images are assigned class labels, and tiles generated from the training images can all be implicitly assigned the class label of the training image, but without assigning a label indicating whether this particular tile contains any feature whose presence means that the image from which the tile was derived belongs to the class indicated in the class label. This training data thus represents weakly annotated data. The learning objective is to train a model that learns to identify features of instances that are highly predictive of the membership of the corresponding bag in a particular class, and thus learns to calculate a prediction value h for each tile for correctly assigning a class label to an image (i.e., to a bag of examples) as a function of the class label predictions generated for each instance of the bag. To generate a class label for the bag from the multiple class label predictions generated by / for all instances of the bag, a permutation invariant pooling function is applied to the predictions generated for each of the bag instances.

[0170] A "MIL program" is a software program that is configured to be or has been trained according to the MIL approach described above.

[0171] A "certainty value" as used herein is a data value, particularly a numerical value, that indicates the certainty of a model of a machine learning program regarding the contribution of a tile's feature vector to the classification of the image from which the tile was derived. For example, if a model predicts that an image belongs to a certain image class based on a pattern observed in the tile, the certainty value indicates the probability that this prediction is correct. The certainty value is calculated, for example, based on a feature vector (including a normalized version of the feature vector, called an "embedding") extracted from one of the tiles of this image. To provide a more specific exemplary explanation, a predictive model may have learned to classify data points that each represent a tile by partitioning the data points in a multidimensional data space with a hyperplane. Data points located far away from the hyperplane may represent a classification result with a "high certainty value," while data points / tiles located very close to the hyperplane may represent a classification result with a low certainty value (a slight change in the attribute / feature value of the data point may place the data point on the other side of the hyperplane and therefore in another class). According to an embodiment of the present invention, the certainty value is represented by a parameter "c".

[0172] A "prediction value" as used herein is a data value, in particular a numerical data value, indicative of the contribution of a tile's feature vector to the classification of the image from which the tile is derived. In particular, the prediction value indicates the contribution of a tile's feature vector to the membership of the image from which the tile is derived in a particular class. For example, the "contribution" can be described as a degree of evidence of the image's membership in a particular class. When calculating the prediction value of a particular tile, the MIL program preferably takes into account only the feature vector of the tile from which the feature vector is derived, and optionally some non-image-based features (but not the feature vectors of other tiles). According to an embodiment of the invention, the prediction value is represented by the parameter "h". A prediction value calculated for a particular tile, for example weighted with the certainty value obtained for this tile, is represented by the parameter "wh". The prediction value can also be described according to an embodiment of the invention as a numerical value indicative of the degree of evidence that the feature vector of the tile provides the membership of the image from which the feature vector of this tile is derived in a particular one of at least two classes.

[0173] As used herein, a "certainty-based pooling function" is a function configured to aggregate feature vectors and / or predictions derived from feature vectors of all tiles of an image into an aggregate prediction, taking into account the certainty calculated for each tile.

[0174] An "aggregate predicted value" as used herein is a data value, e.g., a numerical value, obtained by applying an aggregation function and optionally one or more further data processing steps to a plurality of input values. The input values ​​can be, for example, predicted values ​​(h) or feature vectors (fv) or values ​​derived therefrom. The aggregation function can be, for example, a pooling function configured to aggregate a plurality of feature vectors into a global feature vector (also called an "aggregate feature vector") or to aggregate a plurality of predicted values ​​into an aggregate predicted value, where the pooling function takes into account a certainty value. The one or more further data processing steps can be, for example, the calculation of an aggregate predicted value by a trained machine learning model from the global feature vector, e.g., a trained model of a MIL program. For example, the aggregate predicted value can be a numerical value indicating the likelihood that a particular image belongs to a particular class.

[0175] As used herein, a "heat map" is a graphical representation of data in which individual values ​​contained in a matrix are represented as colors and / or intensity values. According to some embodiments, the heat map is opaque and includes at least some structure of the tissue slide image from which the heat map was generated. According to other embodiments, the heat map is semi-transparent and is displayed as an overlay on top of the tissue image used to generate the heat map. According to some embodiments, the heat map illustrates each of a plurality of similarity scores or a plurality of similarity score ranges via their respective colors or pixel intensities.

[0176] As used herein, a "biomarker-specific dye" is a stain that selectively stains a particular biomarker, e.g., a particular protein such as HER2, or a particular DNA sequence, but generally not other biomarkers or tissue components.

[0177] As used herein, a "non-biomarker specific dye" is a stain that has a more global binding behavior. A non-biomarker specific dye does not selectively stain individual proteins or DNA sequences, but rather stains a larger group of substances with specific physical or chemical properties, as well as subcellular structures, as well as supercellular structures. For example, hematoxylin and eosin are non-biomarker specific dyes, respectively. Hematoxylin is a deep blue or purple stain that is base / positive. It binds to basophilic substances (such as DNA and RNA, which are acidic and negatively charged). Both DNA / RNA in the nucleus and RNA in ribosomes in the rough endoplasmic reticulum are acidic because the phosphate backbones of nucleic acids are negatively charged. These backbones form salts with basic dyes that contain a positive charge. Thus, dyes such as hematoxylin bind to DNA and RNA and stain them purple. Eosin is a red or pink stain that is acidic and negative. It binds to acidophilic substances such as positively charged amino acid side chains (e.g., lysine, arginine). Many proteins in the cytoplasm of some cells are basic because they are positively charged due to arginine and lysine amino acid residues. They form salts with acidic dyes that contain a negative charge, such as eosin. Thus, eosin binds to these amino acids / proteins and stains them pink. This includes cytoplasmic filaments in muscle cells, intracellular membranes, and extracellular fibers.

[0178] The term "intensity information" or "pixel intensity" as used herein is a measure of the amount of electromagnetic radiation ("light") captured on or represented by a pixel of a digital image. The term "intensity information" as used herein can include additional related information, such as the intensity of a particular color channel. A machine learning program, such as a MIL program, can use this information to computationally extract derivative information, such as gradients or textures contained in the digital image. The derivative information can be implicitly or explicitly extracted from the digital image during training and / or during feature extraction by the trained machine learning program. For example, the phrase "pixel intensity values ​​of a digital image correlate with the intensity of one or more particular stains" can imply that the intensity information, including color information, allows a machine learning program, and also a user, to identify areas in a tissue sample that have been stained with a particular one of the one or more stains. For example, a pixel depicting an area of ​​a sample stained with hematoxylin can have a high pixel intensity in the blue channel, and a pixel depicting an area of ​​a sample stained with fastRed can have a high pixel intensity in the red channel.

[0179] The term "biomarker" as used herein is a molecule that can be measured in a biological sample as an indicator of tissue type, normal or pathological processes, or response to a therapeutic intervention. In certain embodiments, the biomarker is selected from the group consisting of proteins, peptides, nucleic acids, lipids, and carbohydrates. More specifically, the biomarker can be a specific protein, such as EGRF, HER2, p53, CD3, CD8, Ki67, etc. Certain markers are properties of specific cells, while other markers are identified as being associated with specific diseases or conditions.

[0180] To determine the stage of a particular tumor based on image analysis of tissue sample images, it may be necessary to stain the sample with multiple biomarker-specific dyes. Biomarker-specific staining of tissue samples typically involves the use of primary antibodies that selectively bind to the biomarkers of interest. In particular, these primary antibodies, along with other components of the staining protocol, may be expensive and therefore may preclude the use of available image analysis techniques for cost reasons in many application scenarios, especially in high-throughput screening. Typically, tissue samples are stained with a background stain ("counterstain"), such as, for example, hematoxylin stain or a combination of hematoxylin and eosin stains ("H&E" stain), to reveal the large-scale tissue morphology and cell-nucleus boundaries. In addition to the background stain, multiple biomarker-specific dyes can be applied depending on the biomedical question that needs to be answered, such as, for example, tumor classification and stage, detection of the amount and relative distribution of specific cell types in a tissue, etc.

[0181] As used herein, a "fully convolutional neural network" is a neural network consisting of convolutional layers without any fully connected layers or multi-layer perceptrons (MLPs) that are typically found at the edges of the network. Fully convolutional nets learn filters in all layers. Even the decision-making layers at the edges of the network learn filters. Fully convolutional nets attempt to learn representations and make decisions based on local spatial inputs.

[0182] According to an embodiment, a full convolutional network is constructed such that its activation function is a vector of output data y at a location (I,j) in a particular layer that satisfies the following property: ij This is a convolutional network that only has layers that generate TIFF0007583041000008.tif8128, where x ij is the data vector at location (i,j) in a particular layer, and y ijis the data vector at that location in the subsequent layer, where y ij is the output generated by the activation function of the network, where k is called the kernel size, s is the stride or subsampling factor, and f ks determines the layer type: matrix multiplication for convolution or average pooling, spatial max for max pooling, or element-wise nonlinearity for activation functions, for other types of layers, etc. This functional form is maintained under composition, with kernel sizes and strides following the following transformation rules: TIFF0007583041000009.tif8128

[0183] While a typical deep net computes a general nonlinear function, a net with only layers of this form computes a nonlinear filter. It is also called a deep filter or fully convolutional network. FCNs naturally handle inputs of any size and produce outputs that correspond to the (possibly resampled) spatial dimensions. For a more detailed description of the properties of various fully convolutional networks, see Jonathan Long, Evan Shelhamer, and Trevor Darrell: "Fully Convolutional Networks for Semantic Segmentation", CVPR 2015.

[0184] As used herein, an "attention machine learning logic program" is an MLL trained to assign weights to certain parameters, where the weights indicate the importance, so that the attention of other programs can be spent on analyzing those parameters. These weights, also called "attention weights," indicate the predictive power of an embedding or tile (but not the certainty of the classification prediction made based on the features of this tile). The idea behind the attention MIL is to simulate the ability of the human brain to selectively pay attention to a subset of available data that is of particular relevance in the current context. The attention MIL is used, for example, in the field of text mining to selectively assign weights and computational resources to certain words that are particularly important for deriving meaning from a sentence. Not all words are equally important. Some of them characterize a sentence more than others. The attention model generated by training the attention MIL on a training dataset can specify that a sentence vector can have more attention to "important" words. According to one embodiment, the trained attention MIL is adapted to calculate a weight for each feature value in each examined feature vector and to calculate a weighted sum of all feature values ​​in each feature vector, which weighted sum embodies the global feature vector of the tile.

[0185] According to an embodiment, the MLL is an attention MIL. The attention MIL is an MLL with a neural attention mechanism adapted to equip the neural network with the ability to focus on a subset of its inputs (or features), computing an attention value a for each input and selecting a particular input based on the attention value. According to an embodiment, let x∈Rd be the input vector, z∈Rk be the feature vector, a∈[0,1]k be the attention vector, g∈Rk be the attention glimpse, and fφ(x) be the attention network with parameter φ. The attention value is given by TIFF0007583041000010.tif4128, where: TIFF0007583041000011.tif4128 is an element-wise multiplication, where z is the output of another neural network fθ(x) with parameter θ. According to an embodiment, the attention values ​​are used as soft attention, which multiplies the features with a (soft) mask of attention values ​​between 0 and 1, or with hard attention if their values ​​are constrained to be exactly 0 or 1, i.e. a∈{0,1}k. In the latter case, the hard attention mask is used to directly index into the feature vector: g~=z[a] (Matlab notation), thereby changing its dimensionality, where g~∈Rm, m≦k. [The present invention 1001] A method (100) for classifying tissue images (212, 712-718, 822-282), comprising: receiving (102) a plurality of digital images (212, 712-718, 822-282) by an image analysis system (200), each of the digital images representing a tissue sample from a patient; - dividing (104) each received image into a set of image tiles (216, 902-906, 940, 942) by the image analysis system; - calculating (106) for each of said tiles a feature vector (220) comprising image features selectively extracted from said tile by said image analysis system; providing a multi-instance learning (MIL) program (226) configured to use a model for classifying any input image as a member of one of at least two different classes based on the feature vectors extracted from all tiles of the input image (108); - calculating (110) for each of said tiles a certainty value (221) indicative of the certainty of said model regarding the contribution of the feature vectors of said tile to the classification of the image from which said tile was derived; -For each image, using a certainty value based pooling function (996) for aggregating (111) the feature vectors extracted from the image by the MIL program (226) into a global feature vector (995) as a function of the certainty values ​​(221) of the tiles of the image, and calculating (112) an aggregated prediction value (997) from the global feature vector; or calculating (113) a prediction value (998) from each of the feature vectors (220) of the image by the MIL program, and using (114) a certainty value based pooling function (996) for aggregating (997) the prediction values ​​for the image into an aggregate prediction value (997) by the MIL program (226) as a function of the certainty values ​​of the tiles of the image; - classifying (116) each of the images as a member of one of the at least two different classes based on the aggregate prediction value by the MIL program; A method comprising: [The present invention 1002] - outputting said classification results to a user via a GUI; and / or - outputting the classification results to another application program; The method of the present invention 1001 further comprises: [The present invention 1003] - the MIL program is a binary MIL program, - said at least two classes consist of a first class, termed the "positive class", and a second class, termed the "negative class", - classifying any one of the images into the "positive class" if the MIL model predicts that, for at least one of the tiles of the image, the feature vector of that tile contains evidence of the "positive class"; - Any one of the images is classified into the "negative class" if the MIL model predicts that for all tiles of that image, their respective feature vectors contain no evidence of the "positive class"; Any of the methods of the present invention. [The present invention 1004] the certainty value-based pooling function is used during testing; Providing the MIL program (108), extracting feature vectors from a set of training tiles generated from a set of training images; - training the MIL program on the feature vectors, thereby using the same certainty value-based pooling function during training as was used during testing, or using a different certainty value-based pooling function during training that is not the certainty value-based pooling function used during testing; Including, Preferably, the certainty value-based pooling function used during training is a certainty value-based max pooling function or a certainty value-based average pooling function, and the certainty value-based pooling function used during testing is a certainty value-based max pooling function. Any of the methods of the present invention. [The present invention 1005] the certainty value based pooling function is a certainty value based max pooling function, and using the certainty value based pooling function comprises, for each of the images, sub-methods a), b), c) or d), each of the sub-methods comprising: - a1) weighting the prediction value (h, 998) of each tile by the certainty value (c, 221) calculated for that tile, thereby obtaining a weighted prediction value (wh, 228); -a2) the maximum of all weighted prediction values ​​calculated for all tiles of the image (wh max ) and -a3) the maximum weighted predicted value (wh max ) as the aggregate predicted value (997); or -b) the maximum certainty value (c max ) as the aggregate prediction (997); or - c1) weighting the feature vector (fv, 220) of each of the tiles by the certainty value (c, 221) calculated for that tile, thereby obtaining a weighted feature vector (wfv); -c2) The maximum of all weighted feature vectors (wfv) calculated for all tiles of the image max ) or -d) the maximum certainty value (c max ) as the global feature vector (995). Any of the methods of the present invention comprising: [The present invention 1006] the certainty value based pooling function is a certainty value based average pooling function; Using the certainty value based pooling function, for each of the images, - weighting the feature vector (fv) of each tile by the certainty value (c) calculated for that tile, thereby obtaining a weighted feature vector (wfv) and calculating the global feature vector as the average of all the weighted feature vectors (wfv) of the image, or - weighting the prediction value (h) of each tile by the certainty value (c) calculated for that tile, thereby obtaining a weighted prediction value (wh), calculating the average of the weighted prediction values ​​for the image, and using the calculated average as the aggregate prediction value. Any of the methods of the present invention comprising: [The present invention 1007] Any of the methods of the invention, wherein the MIL program is a neural network and the certainty value is calculated using a dropout technique when training and / or testing a model of the neural network. [The present invention 1008] The method of the present invention 1007, wherein said certainty value is calculated as a Monte Carlo Dropout (MC Dropout). [The present invention 1009] Any of the methods of the invention, wherein the dropout technique and / or the certainty value based pooling function are used at model testing time rather than at model training time. [The present invention 1010] the neural network includes one or more deactivated dropout layers, the deactivated dropout layers being dropout layers that are activated during training and deactivated during testing, and the method includes reactivating the one or more dropout layers during testing; or the neural network at training time does not include any dropout layers, and the method includes adding one or more dropout layers to the neural network at test time; Calculating the certainty value for any one of the tiles at test time For each of the tiles, a prediction (h d ) multiple times,d each time a different subset of nodes of the network is dropped by the one or more reactivated or added dropout layers; - for each of the tiles, the plurality of predicted values ​​(h d calculating the certainty value for the tile as a function of the variability of (c), the greater the variability, the smaller the certainty value (c); The method of any one of claims 1008 to 1009, further comprising: [The present invention 1011] The received digital image is - digital images of tissue samples, the pixel intensity values ​​of which correlate with the amount of non-biomarker specific stain, in particular hematoxylin stain or H&E stain, and / or - digital images of tissue samples, the pixel intensity values ​​of which correlate with the amount of a biomarker-specific dye adapted to selectively stain the biomarker contained in said tissue sample; Any of the methods of the present invention comprising: [The present invention 1012] providing the MIL program includes training a model of the MIL program; The training, providing a set of digital training images of tissue samples, each digital training image being assigned a class label indicative of one of at least two classes; - dividing each training image into training image tiles, each training tile being assigned the same class label as the digital training image from which it was derived; - calculating, for each of said tiles, a training feature vector comprising image features selectively extracted from said tiles by said image analysis system; and / or - iteratively adapting the model of the MIL program such that an error of a loss function that indicates the difference between the predicted class labels of the training tiles and the class labels actually assigned to the training tiles is minimized, the predicted class labels being calculated by the model based on the feature vectors of the training tiles. Any of the methods of the present invention comprising: [The present invention 1013] For each of said received digital images, - weighting the prediction value (h) of each tile by the certainty value (c) calculated for that tile, thereby obtaining a weighted prediction value (wh); - identifying one of the tiles of the image for which the highest weighted prediction value (wh) is calculated by the MIL program; - calculating, for each other tile of the image, a relevance indicator by comparing the weighted predicted value (wh) of the other tile with the highest weighted predicted value, the relevance indicator being a number that is negatively correlated with the difference between the compared weighted predicted values; - calculating a relevance heatmap (722-728, 822-828) of the image (712-718, 812-828) as a function of the relevance indicators, the pixel colors and / or pixel intensities of the relevance heatmap being indicative of the relevance indicators calculated for the tiles in the image; - displaying the relevance heat map on a GUI; Any of the methods of the present invention further comprising: [The present invention 1014] - displaying the received images on a GUI on a screen, wherein the images are grouped in the GUI into the at least two different classes according to the results of the classification. Any of the methods of the present invention further comprising: [The present invention 1015] Each of the at least two classes is - Patients with sensitivity to certain drugs, - Patients who have developed metastases or certain forms of metastases (e.g., micrometastases), - Cancer patients with a particular response to a particular treatment, e.g., pathological complete response (pCR), - Cancer patient tissues exhibiting specific morphological or microsatellite status, - Patients who have developed adverse reactions to certain medications, - Patients with specific genetic attributes, e.g., a specific genetic signature, and / or - Patients with specific RNA expression profiles Any of the methods of the present invention described above. [The present invention 1016] 13. The method of any of the preceding claims, wherein the certainty value-based pooling function (996) is a certainty value-based max pooling, average pooling, or attention pooling function. [The present invention 1017] - providing a trained attention MLL (222) that learns which tile-derived feature vectors are most relevant for predicting a tile's class membership; - calculating an attention weight (aw) for each of the tiles as a function of the feature vector of the respective tile by the attention MLL, the attention weight being a measure of the relevance of the feature values ​​of that tile with respect to the membership of that tile to a class; - multiplying the attention weight (aw) of the tile by the feature vector value of the tile to obtain an attention-based feature vector having an attention-weighted feature value for the tile; and using the attention-based feature vector as the feature vector input to the MIL program for calculating the predicted value (h), the certainty value (c), and / or the weighted predicted value (wh) of the slide, whereby the predicted value (h), the certainty value (c), and / or the weighted predicted value (wh) are calculated as attention-based predicted value, attention-based certainty value, and / or attention-based weighted predicted value, or - multiplying the attention weight (aw) for the tile by the prediction value (h), the certainty value (c) for the tile, or the weighted prediction value (wh) for the tile calculated by the MIL to obtain the attention-based prediction value, the attention-based certainty value, and / or the attention-based weighted prediction value. Any of the methods of the present invention further comprising: [The present invention 1018] 1. An image analysis system for classifying tissue images, comprising: at least one processor (202); - a volatile or non-volatile storage medium (210) containing digital images each representing a tissue sample from a patient; an image segmentation module (214) executable by said at least one processor and configured to segment each of said images into a set of image tiles (216); a feature extraction module (218) executable by said at least one processor and configured to calculate, for each of said tiles, a feature vector (220) including image features selectively extracted from said tiles; a multi-instance learning (MIL) program (226) executable by the at least one processor and configured to use a model to classify any input image as a member of one of at least two different classes based on the feature vectors extracted from all tiles of the input image; and wherein the MIL program comprises: - calculating (110) for each of said tiles a certainty value indicative of the certainty of said model regarding the contribution of said tile's feature vectors to the classification of the image from which said tile was derived; -For each image, using a certainty value based pooling function (996) for aggregating (111) the feature vectors extracted from the image by the MIL program (226) into a global feature vector (995) as a function of the certainty values ​​(221) of the tiles of the image, and calculating (112) an aggregated prediction value (997) from the global feature vector; or calculating (113) a prediction value (998) from each of the feature vectors (220) of the image by the MIL program, and using (114) a certainty value based pooling function (996) for aggregating (997) the prediction values ​​for the image into an aggregate prediction value (997) by the MIL program (226) as a function of the certainty values ​​of the tiles of the image; - classifying (116) each of the images as a member of one of the at least two different classes based on the aggregate prediction value by the MIL program; The image analysis system further configured for: [Brief description of the drawings]

[0186] In the following, embodiments of the invention will be described in more detail, purely by way of example, with reference to the drawings, in which: [Figure 1] 1 shows a flow chart of a method according to an embodiment of the present invention. [Diagram 2] 1 shows a block diagram of an image analysis system according to an embodiment of the present invention. [Diagram 3] 1 shows a first table showing the comparative test results. [Figure 4] 2 shows a second table showing the comparative test results. [Diagram 5] 1 illustrates the effect of applying dropout to a network layer architecture. [Figure 6]1 illustrates a network architecture of a feature extraction MLL program according to an embodiment of the present invention. [Figure 7] 1 shows a GUI with an image tile gallery. [Figure 8] 1 shows a GUI with a similarity search image tile gallery. [Figure 9] Shows the spatial distance of tiles in 2D and 3D coordinate systems. [Figure 10] 1 illustrates an architecture of a Siamese network according to an embodiment of the present invention. [Figure 11] 1 shows a feature extraction MLL implemented as a truncated Siamese network. [Figure 12] 1 illustrates a computer system that uses feature vector-based similarity searching in an image database. [Figure 13] Shows "similar" and "dissimilar" tile pairs labeled based on the spatial proximity of the tile pairs. [Figure 14] 13 shows feature vectors based on similarity search results extracted by a feature extraction MLL trained on proximity-based similarity labels. [Figure 15] We present an approach I for image classification based on applying a pooling function to the predicted values. [Figure 16] We present approach II for image classification based on applying a pooling function to feature vectors to obtain a global feature vector. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0187] FIG. 1 shows a flow chart of a method according to an embodiment of the present invention. The method can be used to classify tissue images of patients. Classification can be performed, for example, to predict patient attribute values, such as biomarker status, diagnosis, treatment outcome, microsatellite status (MSS) of a particular cancer, such as colorectal or breast cancer, lymph node micrometastasis, and pathological complete response (pCR) in diagnostic biopsies. Prediction is based on classification of digital images of histology slides using a trained MIL program that takes into account model uncertainty. In the following description of FIG. 1, reference is also made to elements of FIG. 2, FIG. 15, and FIG. 16.

[0188] Method 100 can be used to identify previously unknown predictive histological signatures and / or to classify tissue samples with great accuracy.

[0189] In a first step 102, an image analysis system 200 (e.g., as described with reference to FIG. 2) receives a plurality of digital tissue images 212. For example, each tissue image can depict a whole slide tissue sample taken from a patient, e.g., a cancer patient. For each patient in a patient group, at least one digital tissue image is received. For example, the digital tissue images can be read from a local or remote volatile or non-volatile storage medium 210, e.g., a relational database, or can be received directly from an image acquisition device such as a microscope or slide scanner. For example, the received images can each depict a tissue specimen provided, e.g., in the form of an FFPET tissue block slice. At test time, the "endpoint" or "class label" of each received image is unknown, which means that it is unknown whether the patient from whom the tissue sample originates has a particular disease or disease stage, or whether this patient will respond to a particular treatment.

[0190] Next, in step 104, the image analysis system segments each received image into a set of overlapping or non-overlapping image tiles 216. For example, the segmentation can be performed by segmentation module 214.

[0191] According to one embodiment, the received image as well as the generated tiles are multi-channel images.

[0192] Next, in step 106, the image analysis system calculates a feature vector 220 for each of the tiles. The feature vector includes image features selectively extracted from the tissue patterns depicted in the tiles. Optionally, the feature vector may additionally include genetic features or other patient or patient related data available for the patient from which the image and respective tile are derived. According to some embodiments, the feature extraction is performed by a trained feature extraction module 218. The feature extraction module may be a sub-module of the trained MIL program or may be a separate application program or program module. In particular, the feature extraction module 218 may be a trained MLL. The feature extraction MLL may generate a feature vector for each tile while maintaining the feature vector-tile and feature vector-image relationships. Other embodiments may use a feature extraction module implemented as a conventional hand-coded software program that includes feature extraction algorithms to provide a wide variety of features describing the tissue regions depicted in the tiles from which the feature vectors are calculated. Because the features of the feature vector are extracted wholly or at least partially from the image information contained in the respective tile and do not include features extracted from any one of the other tiles of the image, the feature vector represents the optical properties of the tissue region depicted in that tile, and thus may be considered an electronic tissue signature of the tissue depicted in the tile.

[0193] Next, in step 108, a multi-instance learning (MIL) program 226 is provided. For example, the MIL program can be an already trained MIL program that has been installed and / or instantiated in the image analysis system 200, e.g., a computer. The trained and instantiated MIL program then processes tiles of received digital tissue images at test time to classify the received tissue images.

[0194] FIG. 1 illustrates the use of an already trained MIL program to classify tissue images at test time. To generate the trained MIL program provided in step 108, tissue blocks need to be taken from patients with a given known endpoint (e.g., survival, response, gene signature, etc.) and digital training images of these tissue samples are obtained. The endpoints are used as class labels for each training image used to train the MIL program. For example, the tissue blocks are sliced ​​and the slices are set on slides for microscopy. The slices are then stained with one or more histologically relevant stains, such as H&E and / or various biomarker specific stains. Training images are taken from the stained tissue slices using a slide scanner microscope or the like. The training images can be tissue sample images taken many years ago. Old image datasets can have the advantage that they can be used to generate training datasets that include tissue images where the outcomes of many relevant events, such as treatment success, disease progression, side effects, etc., are known at the same time and the known events are assigned as labels. In some embodiments, the training image dataset can include one or more training images per patient. For example, the same tissue sample can be stained multiple times according to different staining protocols, whereby a training image is acquired for each staining protocol. Alternatively, several adjacent tissue sample slices can each be stained using the same or different staining protocols, and a training image is acquired for each tissue sample slide. Each of the received training images is assigned one of at least two different, predefined labels. Each label indicates a patient-related attribute value of the patient whose tissue is depicted in the labeled training image. The attribute values ​​can be of any type, e.g., Boolean, numeric, string, ordinal parameter value, etc. The labels can be assigned to the received training images manually or automatically.For example, a user may configure the slide scanner software so that acquired images are automatically labeled with a particular label during their acquisition. This can be helpful in scenarios where tissue sample images of a large group of patients with the same patient-related attribute values / endpoints are acquired sequentially, such as, for example, 100 tissue images of a first group of 100 breast cancer patients known to respond to a particular drug D, and 120 tissue images of a second group of 120 breast cancer patients known not to respond to this drug. The user may have to set the labels to be assigned to the captured images only once before the training images of the first group are acquired, and then again before the training images of the second group are acquired. Each training image is then divided into multiple tiles, referred to herein as "training tiles." The number of training tiles can be increased by generating modified copies of existing training tiles with different sizes and magnification levels and / or with some simulated artifacts and noise to enhance the training dataset. In some cases, multiple bags can be generated by repeatedly sampling instances in a bag and placing selected instances in additional bags, as described herein for embodiments of the invention. This "sampling" can also have the positive effect of strengthening the training dataset. The MIL program treats each set of tiles of the same image as a bag of tiles for which a bag label needs to be determined. This label determination corresponds to the task of classifying the image, since the image class corresponds to the label for which it is determined.

[0195] The vectors extracted from each digital tissue image received during testing are provided as input to the trained MIL program 226. The feature vectors can be provided in the form of the originally obtained feature vectors or in a normalized form, also called "embeddings."

[0196] Next, in step 110, the MIL program calculates a confidence value 221 for each tile. For example, the model confidence value c for a particular tile k is k a) generate multiple different predictions h using n different dropout layers generated by the dropout technique k1 , h k2 ,...,h kn b) calculating the predicted value h k1 , h k2 ,...,h kn and c) determining the certainty value c as a function of this variability. k The predicted values ​​h calculated based on different dropout network architectures can be calculated by: k1 , h k2 ,...,h kn The high variability of,means that the trained MIL program model is not,“confident” about the contribution of tile-derived features to image,classification results.

[0197] To calculate any of the predictive values ​​h, the MIL program analyzes the feature vectors 220 of the tiles to calculate for each of the tiles a numerical value 228 called the predictive value "h". The predictive value "h" indicates the predictive power of the feature vector associated with the tile for a particular class label potentially assigned to the image. In other words, this numerical value represents the predictive power, or "prognostic value / power", of a particular feature vector for the membership of the image from which the tile was derived in a particular class, given the tissue pattern depicted in this tile. For example, a trained MIL program can learn to predict the contribution of feature values ​​extracted from a tile on the class membership of the image from which the tile was derived. For example, this "contribution" can be a numerical value that correlates with and / or indicates the likelihood that the image belongs to a particular class, given the feature vector of the currently analyzed tile.

[0198] Two approaches to classifying images can be followed.

[0199] According to one approach, in step 113, for each tile of the image to be classified, a prediction value h, 998 is calculated by the MIL program as a function of the feature vector extracted from this tile. This step 113 may already have been performed as part of step 110. However, if it is necessary to calculate a certainty value without using the prediction value h, step 113 needs to be performed. Then, in step 114, a certainty value-based pooling function 996 is applied to the prediction values ​​998 to calculate an aggregate prediction value 997 that integrates the certainty values ​​221 as well as the prediction values ​​calculated for the tiles of the image. This approach according to steps 113 and 114 is shown for example in FIG. 15.

[0200] According to an alternative approach, in step 111, the feature vectors extracted from each tile of the image are input to a certainty value based pooling function 996 which calculates a global feature vector 995 from the feature vectors 220, thereby taking into account the certainty values ​​221 calculated for each tile. The global feature vector integrates the feature vectors calculated for the tiles of the image as well as the certainty values ​​221. Then, in step 112, the global feature vector is input to a predictive model 999 of the MIL program in order to calculate an aggregate predicted value 887. This approach according to steps 111 and 112 is shown for example in FIG.

[0201] Finally, in step 116, each of the digital tissue images 212 received at test is classified according to the aggregate prediction value. For example, if the pooling function is a max pooling function, then if the maximum weighted prediction value wh obtained from all tiles of this image exceeds a predefined threshold, then the image is classified as being a member of a particular "positive" class. Otherwise, the image is classified as not being a member of this particular class. In a binary class setting, this result implies the membership of the image in the other of the two possible classes.

[0202] In some cases, an additional trained attention MIL 222 is provided that has learned which feature vectors are most relevant for predicting class membership. The relevance of a tile is represented as an "attention weight" aw calculated by the attention MIL. In some cases, the attention weight calculated by the attention MIL for each tile is multiplied with the feature vector value of each tile. As a result of the multiplication, a feature vector with weighted feature values ​​is obtained for each tile, which is used as an input to the MIL program to calculate the prediction value h, the certainty value c, and the weighted prediction value wh. In other embodiments, the attention weight aw calculated by the attention MIL is multiplied with the prediction value h or the weighted prediction value wh calculated by the MIL for the feature vector of each tile. This generates a weighted numerical value pp that is used as an indicator of the predictive power of a particular tile and its feature value with respect to class membership.

[0203] According to an embodiment, the trained MIL program and image classification results can be used to stratify patient groups. This means separating patients by factors other than a given treatment. The stratification can be based on patient-related attributes that are not used as labels when training the MIL program. For example, such patient-related attributes can be age, sex, other demographic factors, or specific genetic or physiological traits. The GUI allows the user to select a subgroup of patients whose tissue images were used to train the MIL program based on any one of the patient-related attributes that were not used as labels, and to selectively calculate the accuracy of the trained MLL prediction in that subgroup. For example, the subgroup can consist of female patients, or patients over 60 years old. The accuracy obtained selectively for each subgroup, such as female / male, or patients over / under 60 years old, can reveal a specific high or low accuracy of the trained MIL in some subgroups. This can make it possible to mix in variables (other than the one the researcher is considering), making it easier for the researcher to detect and interpret relationships between the variables and identify patient groups that would benefit most from a particular drug.

[0204] FIG. 2 shows a block diagram of an image analysis system 200 according to an embodiment of the present invention.

[0205] The image analysis system 200 includes one or more processors 202 and a volatile or non-volatile storage medium 210. For example, the storage medium may be a hard disk drive, such as an electromagnetic drive or a flash drive. It may be magnetic storage, semiconductor-based storage, or optical data storage. The storage medium may be a volatile medium, such as a main memory, that contains data only temporarily.

[0206] The storage medium includes a plurality of digital images 212 of tissue samples from a patient with an unknown endpoint.

[0207] The image analysis system comprises a segmentation module 214 configured to segment each of the images 212 into a number of tiles. The tiles are grouped into bags 216, such that typically all tiles in the same bag are derived from the same patient. The bag labels represent the membership of an image in one of two or more different classes and need to be determined and predicted by the trained MIL program 226.

[0208] The feature extraction module 218 is configured to extract a plurality of image features from each of the tiles 216. In some embodiments, the feature extraction module 218 can be a trained MLL or an encoding portion of a trained MLL. The extracted features are stored in the storage medium 210 as a feature vector 220 associated with the tile from which they were derived. Optionally, the feature vector can be augmented with patient features derived from other sources, such as genomic data, such as microarray data.

[0209] The image analysis system includes a multiple instance learning program (MIL program 226). During training, the MIL program 226 receives the feature vectors 220 (or weighted feature vectors generated by considering model uncertainty and optionally also considering attention scores output by the attention MIL 222) as well as the labels assigned to each training tile. As a result of training, a trained MIL program 226 is provided. The trained MIL program has learned to calculate, for each tile, a weighted prediction value wh that indicates the predictive power of the tile and the organizational pattern shown therein for the class label of the image from which the tile was derived, with the weights representing the model certainty (and therefore, implicitly, also the model uncertainty) regarding the class membership predictions calculated based on the feature vector of the particular tile. These weighted prediction values ​​for the tiles of the image are sometimes referred to as "numerical tile relevance scores".

[0210] The image analysis system further comprises a module 230 configured to generate a GUI 232 to be displayed on the screen 204 of the image analysis system.

[0211] The GUI is configured to display the classification results 207 generated by the MIL program for each of the one or more received images 212. For example, the GUI may display an image or image tile with class labels output by the MIL program.

[0212] Optionally, the GUI includes a tile gallery 206 that includes at least some of the tiles and numerical values ​​228 calculated for those tiles. The numerical values ​​228 can be displayed explicitly, for example, as an overlay on each tile, and / or implicitly, for example, in the form of a sort order of the tiles sorted according to their respective numerical values ​​228. When the user selects one of the tiles, a full-slide heatmap of the image from which the tile was originally derived is displayed. In other embodiments, in addition to the tile gallery 206, a heatmap is displayed per default.

[0213] Each of the program modules 214, 215, 218, 222, 226, 230 can be implemented as a submodule of a larger MIL application program or MIL training framework software application. Alternatively, one or more of the modules may each be implemented as a stand-alone software application program that is interoperable with other programs and modules of the image analysis system. Each module and program can be a piece of software written, for example, in Java, Python, C#, or any other suitable programming language.

[0214] Optionally, the image analysis system may comprise a sampling module (not shown) adapted to select a sample (subset) of images for training and testing the trained MIL on the remainder of the image tiles. The sampling module may first cluster the tiles based on their feature vectors before sampling.

[0215] Optionally, the image analysis system can include an attention MIL program 222 configured to calculate weights for each of the feature vectors and each tile in addition to the model certainty based weights. The weights calculated by the attention MIL program can be used together with the feature vectors as input to the MIL program 226 in the training phase and / or to weight the predictions h or weighted predictions wh calculated by the MIL program at test time for each of the tiles.

[0216] 3 and 4 show tables showing comparative test results for using MIL-based image classifiers, where the pooling function considers model uncertainty about the model output calculated for each image tile. An embodiment of the present invention uses a certainty-based pooling function, such as a max pooling function or an average pooling function, to classify tissue images. Tests were performed on several datasets, including a large and challenging real-world dataset (Camelyon16) shown in FIG. 3. The two tests showed that the embodiment of the present invention, especially the use of a certainty-based pooling function, improves the prediction accuracy of most tasks. In the Camelyon16 dataset, the embodiment of the present invention also significantly improved the instance-level prediction accuracy AUC.

[0217] The results further show that by also taking into account the certainty values ​​at training and / or test time, the test-time predictions of differential methods can be significantly improved in most cases. Being able to improve the test-time predictions of existing models is a desirable property. Training MIL methods on large datasets requires expensive infrastructure. Improving test-time of existing models avoids the need to retrain them. Using a certainty value-based max pooling function improves the accuracy of MIL programs at test time.

[0218] FIG. 3 shows a first table 300 showing comparative test results obtained on the dataset used as the basis for the Camelyon-16 lymph node metastasis detection challenge. The goal of this experiment is to compare a MIL program using a certainty-based pooling function with other MIL programs on a challenging real-world dataset. The dataset is available under https: / / camelyon16.grand-challenge.org / data / . Camelyon16 consists of 400 hematoxylin and eosin (H&E) stained whole slide images (WSIs) taken from sentinel lymph nodes that are either healthy or exhibit some form of metastasis. In addition to the WSIs themselves, as well as their labeling (healthy, metastasis-containing), a pixel-level segmentation mask is provided for each WSI. The segmentation masks were discarded and only the global slide-level labels were used to benchmark the MIL programs. A simple staining normalization step was applied to the training images by normalizing the mean and std of the LAB color space channels to be the same as the reference images from the Camelyon16 training set. More advanced normalization can further improve the results. A large set of 256x256 non-overlapping training tiles at 20x resolution (this is the working magnification used by clinical pathologists to review slides) was generated from the training images. The set of training tiles was reduced by automatically identifying and discarding background (white) tiles. Automated background detection of white tiles is described for example in Pierre Courtiol et al.: Classification and Disease Localization in Histopathology Using Only Global Labels: A Weakly-Supervised Approach, 1 February 2018, arXiv:1802.02212. Training tiles are grouped into bags of training tiles, where each bag corresponds to a slide / slide image in the training dataset.A step called staining normalization was applied to normalize tissue staining in the different slides and respective training images.

[0219] A MIL program in the form of Resnet50 (Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition, pp. 770-778, June 2016) pre-trained on ImageNet was provided, and feature vectors (embeddings) of size 2048 were extracted for each tile from the second to the final layer of the network.

[0220] The prediction network used as the MIL program consisted of five fully connected layers with 1024 neurons in the hidden layer, 30% dropout and ReLU activation. When the attention network was applied, four fully connected layers with 1024, 512, and 256 neurons in the hidden layer were used, whereby the network was trained with ReLU, batch normalization, and 30% dropout. In all cases, the Adam optimizer was used with default parameters. During training, 128 random instances were selected from each bag. It was observed that this sampling strategy significantly improved the results and prevented rapid overfitting. For all methods, we selected the model that performed best on the retained validation set (20% of the training set) and tested the selected model on the Camelyon-16 test set. During testing, no example sampling was performed and the entire bag was analyzed. In addition, the ability of the classifier to accurately classify tumor vs. non-tumor tiles was quantified by measuring the instance-level classifier prediction accuracy using expert pathologists' tumor annotations from the Camelyon16 dataset. In addition, for all the methods tested, different pooling functions were applied at testing time to evaluate and compare the performance of the MIL program using the certainty-value-based pooling function method against MIL programs using other pooling strategies when applied to the pre-trained model during inference. The resulting accuracy, measured as "area under the curve - AUC", is shown in Table 1, shown in Figure 3. The certainty-value-based pooling function is called "uncertainty (average or max) pooling" and provides instance-level predictions with the highest accuracy among all the compared MIL programs.

[0221] Figure 4 shows a second table 400 showing the comparative test results obtained on the colon cancer dataset (Sirinukunwattana et al.: “Locality sensitive deep learning for detection and classification of nuclei in routine colon cancer histology Images”, IEEE Transactions on Medical Imaging, 35(5):1196-1206, 2016). This dataset contains 100 H&E images. The dataset was used as a training dataset for the MIL program, where every bag corresponds to a training image and contains 27x27 tiles extracted from the training images. Bags containing malignant regions are given the label 1, otherwise 0. The same network and learning configuration was used as in (MaxiMILian Ilse, Jakub M. Tomczak, Max Welling: “Attention-based Deep Multiple Instance Learning”, 28 June 2018 (v4), ICML 2018 paper, arXiv:1802.04712), except that the learning was not limited to 100 epochs. The training dataset was split into 30% / 70% used for training and testing. 20% of the training set was left as a validation set and the model was selected based on the highest validation score. The results are shown in Table 400. MIL with a certainty value based average pooling function achieved the highest accuracy at both example and bag level.

[0222] In further tests (not shown), the usefulness of the MIL program including the certainty value-based pooling function for key instance retrieval was evaluated based on synthetically generated training and test sets. In this test, 100 bags were created, each with 1,000 instances created from 224 × 224 image tiles. In every bag, 100 instances are set to plain red images. In half of the bags, 100 instances are set to blue and their labels are set to 1. The remaining tiles in each bag are set to be random images from the Camelyon-16 dataset. Images are converted to 2048 embeddings using a Resnet50 pre-trained on ImageNet. Every tile was assigned an attention value using the attention network proposed in Yarin Gal et al., "Deep Bayesian Active Learning with Image Data," March 2017, arXiv:1703.02910v1. Important instances were retrieved from the bags based on these scores, since they influence the contribution of every tile to the output. First, the proposed attention network was trained for 500 epochs on 80% of the bags and tested on the remaining 20% ​​with an AUC of 100%. The instances were ranked using attention scores and certainty value scores from high to low. For the attention network, it was surprisingly observed that the red instances common to all the bags received the highest attention. Red images are not key instances since they do not trigger any labels. Surprisingly, blue images attracted the lowest attention. Thus, in this example, it was observed that the use of attention scores provided suboptimal results. Conversely, certainty value-based scores ranked blue images higher and red images lower. This synthesis problem represents a real-world problem. In histology datasets, uninformative images, mainly containing background or stains, make up a significant portion of the dataset despite preprocessing efforts and can appear in bags of all classes.Training an attention network on such a dataset and retrieving images by their attention scores may result in retrieving uninformative images. This undesirable effect has been observed by the applicant in real digital pathology images. On the other hand, since red images appear in all bags and receive gradients from both labels during training, the model is expected to be uncertain about their labels, and in fact they have the highest uncertainty. Similarly, blue images are associated with only a single label, making the network very certain about them.

[0223] 5 illustrates the effect of applying dropout to a network layer architecture. Network architecture 502 shows the neurons and their connections in three fully connected layers of a neural network. Network 504 shows the same network after applying a dropout layer (mask) to one or more of the layers, thereby deactivating a randomly selected subset of the nodes in each layer, as shown in network architecture 504, where nodes with shading are deactivation nodes.

[0224] Applying a dropout layer during training is a technique (regularization method) that seeks to reduce overfitting (reducing generalization error) by keeping network weights small. More specifically, regularization refers to a class of approaches that add additional information to transform a ill-conditioned problem into a more stable well-conditioned problem. Applying a dropout layer means stochastically removing some nodes and their respective inputs in one or more network layers during training.

[0225] A common problem in machine learning is that a model with too little capacity cannot learn a problem, while a model with too much capacity cannot learn the problem and may overfit the training dataset. In both scenarios, the model does not generalize well. One approach to reduce overfitting when training a neural network is to vary the complexity of the network. The capacity, or complexity, of a neural network model is defined by both the structure of the neural network model in terms of nodes and layers and the parameters in terms of weights. Thus, the complexity of a neural network can be reduced to reduce overfitting by changing the network structure (number of weights) and / or by changing the network parameters (values ​​of the weights). For example, the structure can be adjusted by applying multiple different dropout layers on the fully connected layer of the neural network until an appropriate number of nodes and / or layers is found to reduce or eliminate overfitting for the problem.

[0226] FIG. 6 shows a network architecture 600 of a feature extraction MLL program according to an embodiment of the present invention, which supports a supervised learning approach for generating feature vectors. A deep neural network consisting of a series of autoencoders 604 is trained in a layer-by-layer manner on features extracted from image tiles. The trained network can then perform a classification task, for example classifying tissue depicted in the tile as a member of one of the classes "stromal tissue", "background slide region", "tumor cell", and "metastatic tissue", based on optical features extracted from the image tiles. The network architecture includes a bottleneck layer 606, which has significantly fewer neurons than the input layer 603 and can be followed by further hidden layers and classification layers. According to one example, the bottleneck layer contains approximately 1.5% of the number of neurons of the input layer. Potentially, there are hundreds or thousands of hidden layers between the input layer and the bottleneck layer. The features extracted by the bottleneck layer may be called "deep bottleneck features (DBNF)".

[0227] 7 shows a GUI 700 with an image tile gallery, also referred to as a "report tile gallery." The report gallery (a matrix of tiles under row labels 702, 704, 706, and 708) allows the user to explore the tissue patterns identified by the MIL program as having high predictive power for a particular label. The gallery includes some of the tiles with the highest numerical values ​​for a particular label (or "class") of interest, such as "Response to treatment with drug D=True," as calculated by the MIL program. The tiles are grouped based on the tissue slide image (and respective patient) from which they originated, and sorted within those groups according to their respective predictive values ​​calculated for the class label.

[0228] In the example shown in FIG. 7, a MIL program using attention MIL to weight the prediction values ​​was used to calculate the weights, but the gallery would look similar if the tiles were sorted according to their prediction scores weighted to account for model uncertainty. Thus, according to an embodiment of the present invention, GUI 700 includes a report tile gallery in which tiles are selected and sorted according to their respective calculated weighted prediction values ​​wh, with the weights corresponding to certainty values ​​c. The tile gallery shown in FIG. 7 and FIG. 8 is for illustrative purposes only, since the sort order of tiles is similar to the sort order based on the prediction values ​​or weighted prediction values. Nevertheless, applicants have observed that the accuracy of identifying the most relevant / predictive tiles of a particular class and the accuracy of classification are best when a MIL program including a certainty value-based pooling function is used.

[0229] The report gallery depicted in FIG. 7 may include an overall predictive accuracy for each of the tiles in the gallery, which may be automatically determined after training. Additionally or alternatively, the report gallery may include the label assigned to each image and the bag-wise predictive accuracy obtained for this label. For example, "ground truth=0" may represent the label "patients who respond to drug D" and "ground truth=1" may represent the label "patients who do not respond to drug D". The highest value of all tiles of a particular image calculated by the MIL is displayed as a "prediction value" h or a "weighted prediction value" wh on the group of tiles derived from said image.

[0230] In the depicted gallery, tile row 702 shows six tiles for a first patient. The first one of the tiles is assigned the highest predictive (prognostic) value indicating the predictive power of the particular tissue slide / whole slide image with respect to the label. The first tile per slide group is additionally or alternatively assigned the highest combined value (derived from the numerical values ​​provided by the MIL and from the weights calculated by the attention MIL) of all tiles derived from the particular tissue slide image.

[0231] The highest predictive values ​​can be displayed at the top of the highest scoring tiles for each patient, as depicted in the GUI shown in FIG.

[0232] A report tile gallery that includes only the subset of tiles with the highest predictive value or highest weighted predictive value can be advantageous because it does not require the pathologist to inspect the entire slide. Rather, the pathologist's attention is automatically directed to a small number of subregions (tiles) of each full slide image whose tissue patterns have been identified as having the highest predictive power for the label of interest.

[0233] According to the embodiment depicted in Figure 7, the report image tile gallery shows image tiles derived from H&E stained images. The report image tile gallery is organized as follows:

[0234] Row 702 contains the six tiles that have been assigned the highest numerical values ​​(indicative of predictive power, i.e. prognostic value) calculated by the attention-based MIL program among all tiles derived from a particular full slide image 712 of the first patient. According to other embodiments, the sorting is performed based on a score value that is identical to the predictive value calculated by the MIL or is a derivative of the numerical value calculated by the MIL, e.g., a weighted predictive value wh.

[0235] Each full slide image 712 of the first patient's tissue sample that was used to generate the tile, some of which are presented in row 712, is shown in spatial proximity to this selected set 712 of more highly relevant tiles.

[0236] In addition, an optional relevance heatmap 722 is shown that highlights all full slide image regions whose weighted prediction values ​​calculated by the MIL are similar to the value of one of the tiles of the image 712 for which the highest value indicative of predictive power was calculated. In this case, one of the tiles with the highest value calculated is automatically identified and selected (e.g., the tile in the first position in row 712) and used as the basis for calculating the relevance heatmap 722. According to an alternative implementation, the relevance heatmap 722 does not represent the similarity of the tile's value to the highest value calculated for all tiles of the image, but rather to the highest combined score calculated for all tiles of the image. The combined score can be a combination, such as, for example, a multiplication of the weight calculated by the attention MIL for the tile and a value indicative of the tile's predictive power with respect to the label of the image, calculated by the MIL. According to a still further embodiment, the relevance heatmap 722 represents the similarity of the tile weights calculated by the attention MIL to the highest weights calculated by the attention MIL for all tiles of the image.

[0237] Column 704 includes the six tiles that were assigned the highest numerical values ​​calculated by the MIL program within all tiles derived from a particular whole slide image 714 of the second patient. Each whole slide image 714 is shown in spatial proximity to this selected set of more highly relevant tiles. In addition, a relevance heat map 724 is shown that highlights all whole slide image regions whose respective numerical values ​​calculated by the MIL are highly similar to one of the tiles of the whole slide image 714 with the highest numerical value calculated by the MIL.

[0238] Column 706 includes the six tiles that were assigned the highest numerical values ​​calculated by the MIL program within all tiles derived from a particular whole slide image 716 of the third patient. Each whole slide image 716 is shown in spatial proximity to this selected set of more highly relevant tiles. In addition, a relevance heat map 726 is shown that highlights all whole slide image regions whose respective numerical values ​​calculated by the MIL are highly similar to one of the tiles of the whole slide image 716 whose highest numerical value was calculated by the MIL.

[0239] Column 708 contains the six tiles that have been assigned the highest numerical values ​​calculated by the MIL program within all tiles derived from a particular whole slide image 718 for a patient. Each whole slide image 718 is shown in spatial proximity to this selected set of more highly relevant tiles. In addition, a relevance heat map 728 is shown that highlights all whole slide image regions whose respective numerical values ​​calculated by the MIL are highly similar to one of the tiles of the whole slide image 718 with the highest numerical value calculated by the MIL.

[0240] According to an embodiment, the relevance heatmap presented in the report tile gallery shows predictive power, or attention-based weighting, and / or certainty-based weighting, or a combination thereof. In the depicted example, bright pixels in the heatmap depict areas in the image where tiles have high predictive power, high certainty-based weighting, or a combination thereof. According to an embodiment, calculating the relevance heatmap includes determining whether the tile's score (e.g., numerical, weighted, or combined value) exceeds a minimum percentage value of the scores of the highest-scoring tiles of the image. If this is exceeded, each tile in the relevance heatmap is represented by a first color or a "bright" intensity value, e.g., "255". If this is not exceeded, each tile in the relevance heatmap is represented by a second color or a "dark" intensity value, e.g., "0".

[0241] Each tile in the report tile gallery can be selected by the user to initiate a similarity search (e.g., by double-clicking on the tile or by selecting the tile with a single click and then selecting the GUI element "Search"), followed by displaying a similarity search tile gallery, e.g., as shown in FIG.

[0242] The "Blacklist" and "Retrain" elements in the set of selectable GUI elements 710 allow the user to define a blacklist of tiles and have the MIL program retrain on all tiles except those in the blacklist and those highly similar to the tiles in the blacklist. For example, the blacklist can include a set of manually selected tiles that have a particularly low numerical value (prognostic value) or a particularly high numerical value due to their inclusion of artifacts (excluding tiles with very high predictive power can increase the ability of the MIL to identify additional, previously unknown tissue patterns that also have predictive power for the label of interest). In response to a user adding a particular tile to the blacklist, the image analysis system can be configured to automatically identify all tiles whose feature vector-based similarity to the feature vector of the blacklisted tile exceeds a minimum similarity threshold. The identified tiles are likewise automatically added to the blacklist. When the user selects the retrain GUI element, the MIL is retrained on all tiles of the training data set except those in the blacklist.

[0243] 8 illustrates a GUI 800 with a similarity search image tile gallery according to an embodiment of the present invention. The similarity search is triggered by a user-based selection of one of the tiles 830 in the report gallery 700.

[0244] This search identifies a subset of, for example, the six most similar tiles based on the similarity of the compared feature vectors within tiles generated from each of the full slide images 812 to 818. The tiles identified in the similarity search are grouped by full slide image or by patient and sorted in descending order according to their similarity to the tile 830 (the "query tile"), whose selection triggered the similarity search.

[0245] The whole slide images 812-818 and similarity heatmaps 822-828 indicate the locations of tiles whose feature vectors (and therefore the depicted tissue patterns) are most similar to the feature vector of the selected tile.

[0246] Optionally, the similarity search tile gallery additionally includes one or more of the following data: - The label assigned to the image from which the depicted tile was derived. One label depicted in Figure 8 is “Ground Truth: 0”, - the predicted accuracy calculated by the MIL program for each bag (image) with respect to the label on the bag, - the number of similar tiles in the entire slide image and / or the percentage of similar tiles compared to non-similar ones (e.g., by a threshold); and -The mean, median or histogram of the similarity values ​​of all tiles across all slide images.

[0247] For example, the user may select one of the tiles 830 in the report gallery that assigned the highest predictive value (e.g., highest h or wh) and therefore the highest predictive power for the image's label / class membership. By selecting a tile, the user may initiate a tile-based similarity search across tiles and images of many different patients who may be assigned a different label than the currently selected tile. The similarity search is based on a comparison of feature vectors and tiles to determine similar tiles and similar tissue patterns based on similar feature vectors. These are by assessing and displaying the number and / or percentage of tiles (and respective tissue patterns) that are similar to the selected tile (and its tissue pattern) but have a different label than the selected tile's label (e.g., "Reaction to Drug D = Fake Patient" instead of "Reaction to Drug D = True Patient").

[0248] Thus, by selecting a tile returned by the MIL program as "highly predictive," a pathologist can easily check the predictive power, especially the sensitivity and specificity, of the tissue pattern identified by the MIL program and perform a similarity search that reveals how many tiles in the data set with similar feature vectors are assigned the same label as the selected tile. This is a major advantage over state-of-the-art machine learning applications that can also provide an indication of predictive features of tissue images, but do not allow the user to identify or verify these features. Based on the report gallery and similarity search gallery, a human user can verify the proposed highly predictive tissue patterns and can also verbalize the common features and structures that are shown in all tiles that have high predictive power and are associated with similar feature vectors.

[0249] The feature that tiles in the report gallery are selectable and the selection triggers the execution of a similarity search to identify and display other tiles with similar feature vectors / tissue patterns can allow the user to freely select any image tile in the report tile gallery that is of interest to the user. For example, a pathologist can be interested in tissue patterns and respective tiles with the highest predictive power (highest numerical value calculated by MIL) as described above. Alternatively, a pathologist may typically be interested in artifacts with a certain low predictive power (certain low numerical value). Further alternatively, a pathologist may be interested in a particular tissue pattern for any other reason, for example because it reveals some side effect of a drug or any other relevant biomedical information. A pathologist freely selects any one of the tiles in the respective report tile gallery. This triggers the pathologist to perform a similarity search and to calculate and display the results in the form of a similarity tile gallery. This display and GUI can be automatically refreshed after the completion of the similarity search.

[0250] According to some embodiments, the computation and display of the similarity search gallery includes the computation and display of a similarity heat map 822-828. The heat map encodes similar tiles and their respective feature vectors in color and / or pixel intensity. Image regions and tiles with similar feature vectors are represented in the heat map with similar color and / or high or low pixel intensity. Thus, the user can quickly obtain an overview of the distribution of a particular tissue pattern signature in the entire slide image. The heat map can be easily refreshed by simply selecting a different tile, since this selection automatically triggers a recalculation of the similarity of the feature vectors based on the feature vectors of the newly selected tile.

[0251] According to an embodiment, the affinity search gallery includes an affinity heat map. The method includes generating the affinity heat map by a subordinate method. The subordinate method includes: -Selecting one of the tiles in the report tile gallery; - for each of some or all of the other tiles of the received image, calculating, with respect to the selected tile, a similarity score by comparing the feature vectors of the other tiles of the same image and other images with the feature vector of the selected tile; - calculating, for each of the images whose tiles have been used to calculate a respective similarity score, a respective similarity heatmap as a function of the similarity score, pixel color and / or pixel intensity of the similarity heatmap indicating the similarity of the tile in said image to the selected tile; - displaying a similarity heatmap.

[0252] According to an embodiment, the image tiles shown in the similarity search gallery are also selectable.

[0253] The similarity heatmap can provide useful overview information that allows a human user to easily recognize how widespread a particular tissue pattern of interest is in a particular tissue or in tissue samples of a subgroup of patients with a particular label. The user can freely select any tile in the search gallery, which triggers a recalculation of the similarity heatmap based on the feature vector assigned to the currently selected tile and an automatic refresh of the GUI containing the similarity heatmap, respectively.

[0254] According to an embodiment, the image tiles in the report gallery and / or in the similarity search tile gallery are grouped based on the patients from whose tissue sample images the tiles are derived. According to an alternative embodiment, the image tiles in the report gallery and / or in the similarity search tile gallery are grouped based on the labels assigned to the images from which the tiles are derived.

[0255] Typically, all images derived from the same patient will have the same label, and all tiles derived from those images of a particular patient will be treated by MIL as members of the same "bag". However, in some exceptional cases, different images of the same patient may be assigned different labels. For example, if a first image depicts a first metastasis in a patient and a second image depicts a second metastasis in the same patient, and it is observed that in response to treatment with drug D, the first metastasis disappears while the second metastasis continues to grow, then patient-related attribute values ​​may be assigned per image instead of per patient. In this case, there may be multiple bags of tiles per patient.

[0256] According to another example, images of a patient's tissue sample are acquired before and after treatment with a particular drug, and the endpoint (label) used to train the MIL program and / or to apply the trained MIL program is the attribute value "tissue state=post-treatment with drug D" or the attribute value "tissue state=pre-treatment with drug D". Training the MIL program on said patient-related attribute values ​​can have the advantage of identifying tissue patterns indicative of the drug's activity and morphological effects on the tumor. Such identified tissue patterns of drug effects can allow for validating and exploring the drug's mode of action as well as potential adverse drug effects.

[0257] According to an embodiment, the method further includes computationally increasing the number of bags of tiles by generating additional sets of tiles, each additional set of tiles being treated by the MIL program as an additional bag of tiles that is assigned the same label as the tissue image from which the saw tile was generated. The generation of the additional sets of tiles includes, inter alia, applying one or more artifact generation algorithms to at least one subset of tiles to generate new tiles that include artifacts. Additionally or alternatively, the generation of the additional bags of tiles can include increasing or decreasing the resolution of at least one subset of tiles to generate new tiles that are finer or coarser than their respective saw tiles.

[0258] For example, the subsets can be obtained for each patient by randomly selecting some or all tiles of one or more tissue images obtained from said patient. The artifact generation algorithm simulates image artifacts. The image artifacts can be, for example, artifacts of types generated during tissue preparation, staining, and / or image acquisition (e.g., edge artifacts, over-staining, under-staining, dust, speckle artifacts, (simulated by Gaussian blur, etc.)). Additionally or alternatively, the artifacts can be generic noise types (e.g., simulated by occlusion, color jittering, Gaussian noise, salt & pepper, rotation, flip, skew distortion, etc.).

[0259] The generation of additional bags of tiles can have the advantage that additional training data is generated from a limited set of available training data. The additional training data represents image data that may be degraded by common distortions, artifacts, and noise that often occur in the context of sample preparation and image acquisition. Thus, the expanded training data set can ensure that overfitting of the model, which is inherent in MIL programs during training, is avoided.

[0260] According to an embodiment, the method further comprises calculating clusters of tiles taken from one or more received digital images, the tiles being grouped into clusters based on the similarity of their feature vectors. Preferably, a cluster is calculated for each of the patients. This means that tiles from different images depicting different tissue slides of the same patient can be grouped into the same cluster if their feature vectors are sufficiently similar.

[0261] According to another embodiment, clusters are computed jointly for all tiles from all patients.

[0262] In both clustering methods (all tiles from different patients together or per patient), tiles that look similar to each other (i.e. have similar feature vectors) are clustered into the same cluster.

[0263] For example, in the "cluster all tiles for different patients" case, the clustering result may be the generation of 64 groups (clusters) of tiles for all tiles for all patients, each of which contains similar tiles derived from a different patient. Conversely, in the case of clustering by patient, each patient will have its own 64 clusters.

[0264] When clusters are generated per patient, there may be cases where the patient's images have no tiles containing fat or tiles with very little fat. In this case, a "fat cluster" may not be generated because there is not enough data to learn a cluster around the feature vector that characterizes that "fat". However, performing a method of clustering all tiles of all patients together may have the advantage that a large number of clusters / tissue types can be identified with the largest amount of available data. With "all patient tiles" clustering, it is more likely that a cluster for "fat" tissue patterns will be identified since at least some patients will have some fat cells in their biopsies. Thus, if the number of tiles depicting fat cells in the dataset is likely to be sufficient, a cluster for fat cells will be generated (even for patients with very low fat cell content). If clusters are generated for all tiles of all patients together and one cluster represents fat cells, then all tiles with fat cells from all patients are grouped into that cluster. This means that for a particular patient / bag, all tiles with fat cells are grouped together in the cluster, and when cluster sampling is used for that bag, some of the tiles (from the current patient / bag) that belong to the cluster are selected.

[0265] Clustering tiles may be advantageous because this operation can reveal the number and / or type of tissue patterns observable in a particular patient. According to some embodiments, the GUI includes a user-selectable element that allows a user to trigger the clustering of tiles and the presentation of the tile clusters in a clustered gallery view. This can assist a user in intuitively and quickly understanding important types of tissue patterns observed in a particular tissue sample from a patient.

[0266] According to an embodiment, training the MIL program includes repeatedly sampling a set of tiles to extract a subset of tiles from the set of tiles, and training the MIL program on the subset of tiles.

[0267] As used herein, the term "sampling" is a technique used in the context of data analysis or training machine learning algorithms that involves taking a specifically chosen number of L samples (here, instances, i.e. tiles) from a number of N data items (instances, tiles) in a dataset (a sum of tiles taken from one or more images of a patient). According to an embodiment, "sampling" involves selecting a subset of data items from within the number of N data items according to a probability distribution that is estimated to be statistically representative of the sum of the N tiles in the training dataset. This can allow for more accurate learning of the properties of the entire population. The probability distribution represents the statistical assumptions that guide the machine learning process and make "learning from data" feasible.

[0268] According to some embodiments, sampling is performed by randomly selecting a subset of tiles to provide a bag of sampled tiles.

[0269] According to an embodiment, clustering and sampling are combined as follows: sampling involves selecting tiles from each of the tile clusters acquired for the patient such that the number of tiles in each tile subset generated in sampling corresponds to the size of the cluster from which said tile was acquired.

[0270] For example, 1000 tiles can be generated from a digital tissue image of a particular patient. Clustering produces a first cluster containing 300 tiles representing background tissue slide regions, a second cluster containing 400 tiles representing stromal tissue regions, a third cluster containing 200 tiles representing metastatic tumor tissue, a fourth cluster containing 40 tiles representing specific staining artifacts, and a fifth cluster containing 60 tiles representing tissue with microvessels.

[0271] According to one embodiment, sampling involves selecting a certain percentage of tiles from each of the clusters, such as 50%, which would mean 150 tiles from cluster 1, 200 tiles from cluster 2, 100 tiles from cluster 3, 20 tiles from cluster 4, and 30 tiles from cluster 5.

[0272] According to a preferred embodiment, sampling involves selecting an equal number of tiles from each cluster. This sampling approach can have the advantage that the same number of tile / tissue pattern examples are drawn from different types of clusters, which can make the training dataset more balanced. This can make the trained MIL and / or the trained attention MIL more accurate when the desired predictive features are rare in the training dataset.

[0273] The combination of clustering and sampling can be particularly advantageous because it can increase the data basis for training without unintentionally "losing" some tiles that are actually of high predictive power through sampling. In the context of digital pathology, often the vast majority of areas of a tissue sample are altered by a particular disease or other patient-related attributes and do not contain tissue regions that are predictive. For example, only a small sub-region of a tissue sample may actually contain tumor cells, while the rest may represent normal tissue. Clustering tiles first and then selecting tiles from each of the clusters can ensure that at least some of the small amount of tiles that represent predictive tissue patterns, such as tumor cells or microvessels, always become part of the sample.

[0274] Figure 9 shows the spatial distance of tiles in 2D and 3D coordinate systems, which is used to automatically assign similarity labels to pairs of tiles based on similarity labels derived automatically from the spatial proximity of the tiles. This provides a training dataset to train the feature extraction MLL without requiring manual annotation of images or tiles by domain experts.

[0275] FIG. 9A illustrates the spatial distance of tiles in a 2D coordinate system defined by the x- and y-axes of a digital tissue sample training image 900. The training image 900 depicts a tissue sample of a patient. After the tissue sample is obtained from the patient, the sample is placed on a microscopy slide and stained with one or more histologically relevant stains, e.g., H&E, and / or various biomarker-specific stains. The training image 900 is captured from the stained tissue sample using a slide scanner microscope or the like. According to some implementation variants, at least some of the received training images are derived from different patients and / or are derived from different tissue regions (biopsies) of the same patient and therefore cannot be aligned with each other in a 3D coordinate system. In this case, the tile distance can be calculated in the 2D space defined by the x- and y-coordinates of the image, as described below.

[0276] A training image 900 is divided into tiles. For purposes of illustration, the size of the tiles in FIG. 9A is larger than a typical tile size.

[0277] The training dataset can be automatically labeled by the following approach: First, a starting tile 902 is selected. Then, a first circle area around this starting tile is determined. The radius of the first circle is also referred to as a first spatial proximity threshold 908. All tiles within this first circle, e.g., tile 906, are considered to be "near" tiles to the starting tile 902. In addition, a second circle area around this starting tile is determined. The radius of the second circle is also referred to as a second spatial proximity threshold 910. All tiles outside this second circle, e.g., tile 904, are "far" tiles with respect to the starting tile 902.

[0278] A first set of tile pairs is then generated, with each tile pair in the first set including a starting tile and a "nearby" tile of the starting tile. For example, this step may include generating as many tile pairs as there are nearby tiles included in the first circle. Alternatively, this step may include randomly selecting a subset of the available nearby tiles and generating a tile pair for each of the selected nearby tiles by adding the starting tile to the selected nearby tile.

[0279] A second set of tile pairs is generated, each tile pair of the second set including a starting tile and a "far" tile with respect to the starting tile. For example, this step may include generating as many tile pairs as there are far tiles outside the second circle included in the image 800. Alternatively, this step may include randomly selecting a subset of the available far tiles and generating a tile pair for each of the selected far tiles by adding the starting tile to the selected far tile.

[0280] Then, another tile in the image 900 can be used as the starting tile, and the above steps can be performed similarly. This means that the first and second circles are drawn again using the new starting tile as the center. This identifies nearby tiles and far tiles with respect to the new starting tile. The first set of tiles is populated with the pair of nearby tiles identified based on the new starting tile, and the second set of tiles is populated with the pair of far tiles identified based on the new starting tile.

[0281] Subsequently, yet another tile in the image 900 can be selected as the starting tile and the above steps can be repeated, thereby further adding additional tile pairs to the first and second tile pair sets. Selection of new starting tiles can be performed until all tiles in the image have been selected once as starting tiles, or until a predetermined number of tiles have been selected as starting tiles.

[0282] To each of the tile pairs in the first set, e.g., pair 912, the label "similar" is assigned. To each of the tile pairs in the second set, e.g., pair 914, the label "dissimilar" is assigned.

[0283] FIG. 9B shows the spatial distance of tiles in a 3D coordinate system defined by the x-axis and y-axis of the digital tissue sample image 900 and the z-axis corresponding to the bundle height of the images 900, 932, 934 aligned with each other according to the relative positions of the tissue slices of the tissue block depicted by the training images 900, 932, 934, respectively. The training images each depict a tissue sample derived from a single tissue block of a particular patient. The depicted tissue sample belongs to a bundle of multiple adjacent tissue slices. For example, this bundle of tissue slices can be prepared ex vivo from a FFPET tissue block. The tissue block is sliced ​​and the slices are set on a slide for microscopy. The slices are then stained as described for image 900 with reference to FIG. 8A.

[0284] Since the tissue samples in this bundle are derived from a single tissue block, the digital images 900, 932, 934 can be aligned in a common 3D coordinate system, whereby the z-axis is orthogonal to the tissue slice. The z-axis is the axis orthogonal to the tissue slice. The distance of the images in the z-direction corresponds to the distance of the tissue slice depicted by the image. If the two tiles of a pair are derived from the same image, the tile distance of the tile pair is calculated in 2D space. In addition, tile pairs can be generated, whose tiles are derived from different images that are aligned with each other in a common 3D coordinate system. In this case, the distance of the two tiles in the pair is calculated using the 3D coordinate system.

[0285] Each of the aligned digital images is divided into multiple tiles. For purposes of illustration, the tile size in FIG. 9B is larger than a typical tile size.

[0286] The training dataset can be automatically labeled by the following approach: First, a starting tile 902 is selected. Then, tile pairs including the starting tile and a nearby tile and tile pairs including the starting tile and a far tile are identified and labeled as described below.

[0287] A first 3D sphere is determined around this starting tile. For purposes of illustration, only a cross section of the first sphere is shown. The radius of the first sphere is also referred to as a first spatial proximity threshold 936. All tiles within this first sphere are considered to be "near" tiles of the starting tile 902, for example, tile 906 in image 900 as well as tile 940 in image 934. In addition, a second sphere is determined around this starting tile. The radius of the second sphere is also referred to as a second spatial proximity threshold 938. All tiles outside this second sphere are "far" tiles with respect to the starting tile 902, for example, tile 904 in image 900 as well as tile 942 in image 934.

[0288] A first set of tile pairs is generated, each tile pair in the first set including a starting tile and a "nearby" tile of the starting tile. For example, this step may include generating as many tile pairs as there are nearby tiles included in the first sphere. Alternatively, this step may include randomly selecting a subset of the available nearby tiles and generating a tile pair for each of the selected nearby tiles by adding the starting tile to the selected nearby tile.

[0289] A second set of tile pairs is generated, each tile pair of the second set including a starting tile and a "far" tile with respect to the starting tile. For example, this step may include generating as many tile pairs as there are far tiles included in the images 900, 932, 934 outside the second sphere. Alternatively, this step may include randomly selecting a subset of the available far tiles and generating a tile pair for each of the selected far tiles by adding the starting tile to the selected far tile.

[0290] Then, another tile in image 900 or in images 932, 934 can be used as the starting tile, and the above steps can be performed similarly. This means that the first and second spheres are redrawn using the new starting tile as the center. This identifies nearby tiles and far tiles with respect to the new starting tile. The first set of tiles is populated with the pair of nearby tiles identified based on the new starting tile, and the second set of tiles is populated with the pair of far tiles identified based on the new starting tile.

[0291] The above steps can be repeated until all tiles in each of the received images 900, 932, 934 have been selected as starting tiles (or until another termination criterion is met), thereby adding further tile pairs to the first and second tile pair sets.

[0292] To each of the tile pairs in the first set, e.g., pairs 912 and 913, the label "similar" is assigned. To each of the tile pairs in the second set, e.g., pairs 914 and 915, the label "dissimilar" is assigned.

[0293] The circle and sphere based distance calculations shown in Figures 9A and 9B are just examples for calculating distance based similarity labels, in this case binary labels of either "similar" or "dissimilar". Other approaches could be used, such as, for example, calculating the Euclidean distance between two tiles in a 2D or 3D coordinate system, and calculating a numerical similarity value that is negatively correlated with the Euclidean distance of the two tiles.

[0294] Because the number of pixels corresponding to 1 mm of tissue depends on various factors such as the magnification of the image capture device and the resolution of the digital image, all distance thresholds herein are specified with respect to the actual physical object being depicted, i.e., the tissue sample or a slide covered with a tissue sample.

[0295] 10 shows the architecture of a Siamese network trained according to an embodiment of the present invention to provide a sub-network capable of extracting biomedically meaningful feature vectors from image tiles suitable for performing feature vector-based similarity searches and / or feature vector-based clustering of tiles. The Siamese network 1000 is thus trained on an automatically labeled training dataset including tile pairs with automatically generated proximity-based similarity labels, e.g., as described with reference to FIG. 9A and / or FIG. 9B.

[0296] The Siamese network 1000 consists of two identical sub-networks 1002, 1003 connected at their output layer 1024. Each network includes an input layer 1005, 1015 adapted to receive as input a single digital image (e.g., a tile) 954, 914. Each of the sub-networks includes several hidden layers 1006, 1016, 1008, 1018. A one-dimensional feature vector 1010, 1020 is extracted by each one of the two sub-networks from one of the two input images. The last hidden layer 1008, 1018 of each network is then adapted to calculate a feature vector and provide it to the output layer 1024. The processing of the input images is strictly separated. This means that the sub-network processes only the input image 1054 and the sub-network processes only the input image 1014. The only point at which the information provided in the two input images is combined is in the output layer, where it compares two vectors to determine vector similarity, and therefore the similarity of the tissue patterns depicted in the two input images.

[0297] According to an embodiment, each sub-network 1002, 1003 is based on a modified resnet-50 architecture (He et al., Deep Residual Learning for Image Recognition, 2015, CVPR'15). According to an embodiment, the resnet-50 pre-trained sub-networks 902, 903 are pre-trained on ImageNet. The last layer (which typically outputs 1,000 features) is replaced with a fully connected layer 1008, 1018 of a size having a desired size of the feature vector, for example size 128. For example, the last layer 1008, 1018 of each sub-network can be configured to extract features from the penultimate layer, such that the penultimate layer can provide a larger number of features (e.g., 2048) than the last layer 1008, 1018. According to an embodiment, an optimizer such as, for example, the Adam optimizer with default parameters in PyTorch (learning rate of 0.001 and beta of 0.9, 0.999) and a batch size of 256 was used during training. For data augmentation, random horizontal and vertical flips and / or random rotations up to 20 degrees and / or color jitter expansion with values ​​of 0.075 for luminance, contrast saturation and / or hue can be applied to the tiles to augment the training dataset.

[0298] When a Siamese network is trained on pairs of automatically labeled images, the goal of the learning process is that similar images should have outputs (feature vectors) that are similar to each other, and dissimilar images should have outputs that are dissimilar to each other. This can be achieved by minimizing a loss function, such as a function that measures the difference between the feature vectors extracted by the two sub-networks.

[0299] According to an embodiment, a Siamese neural network is trained for pairs of tiles using a loss function such that the similarity of the feature vectors extracted by the two sub-networks for each of the two tiles of the pair correlates with the similarity of the tissue patterns depicted in the two tiles of the pair.

[0300] The Siamese network can be, for example, a Siamese network as described in Bromley et al., Signature Verification using a 'Siamese' Time Delay Neural Network, 1994, NIPS'1994. Each sub-network of the Siamese network is adapted to extract a multi-dimensional feature vector from each one of the two image tiles provided as input. The network is trained on multiple tile pairs that have been automatically annotated with proximity-based tissue pattern similarity labels with the goal that tile pairs depicting similar tissue patterns should have outputs (feature vectors) that are close to each other (similar) and tile pairs depicting dissimilar tissue patterns should have outputs that are far from each other. According to one embodiment, this is achieved by performing a contrastive loss, for example, as described in Hadsell et al., Dimensionality Reduction by Learning an Invariant Mapping, 2006, CVPR'06. The contrastive loss is minimized during training. The contrastive loss CL can be calculated, for example, according to: TIFF0007583041000012.tif7128 where f1, f2 output two identical sub-networks and y is the ground truth label of the tile pairs: 0 if they are labeled as "similar" (first set of tile pairs) and 1 if they are labeled as "dissimilar" (second set of tile pairs).

[0301] Training the Siamese network 1000 involves feeding the network 1000 with a number of automatically labeled similar tile pairs 912, 913 and dissimilar tile pairs 914, 915. Each input training data record 1028 includes two tiles of a tile pair and their automatically assigned spatial proximity-based labels 1007. The proximity-based labels 10007 are provided as "ground truth". The output layer 1024 is adapted to calculate predicted similarity labels for the two input images 1004, 1014 as a function of the similarity of the two compared feature vectors 1008, 1018. Training the Siamese network involves a backpropagation process. Any deviation of the predicted label 926 from the input label 1007 is considered an "error" or "loss" measured in the form of a loss function. Training the Siamese network involves minimizing the error calculated by the loss function by iteratively using backpropagation. The Siamese network 1000 can be implemented, for example, as described by Bromley et al., "Signature Verification using a ''Siamese'' Time Delay Neural Network," 1994, NIPS'1994.

[0302] FIG. 11 illustrates a feature extraction MLL 950 implemented as a truncated Siamese network, for example as described with reference to FIG.

[0303] The feature extraction MLL 950 can be obtained, for example, by separately storing one of the sub-networks 1002, 1003 of the trained Siamese network 1000. In contrast to the trained Siamese network, the sub-network 1002, 1003 used as feature extraction MLL requires only a single image 952 as input and does not output a similarity label, but rather outputs a feature vector 954 selectively including values ​​of a limited set of features identified during the training of the Siamese network 1000 as being particularly characteristic of a particular tissue pattern and particularly suitable for determining the similarity of tissue patterns depicted in two images by extracting and comparing this particular set of features from the two images.

[0304] Figure 12 shows a computer system 980 that uses a similarity search based on feature vectors in an image database. For example, the similarity search can be used to compute a search tile gallery, an example of which is depicted in Figure 8. The computer system 980 comprises one or more processors 982 and a trained feature extraction MLL 950, which can be a sub-network of a trained Siamese network ("truncated Siamese network"). The system 980 is adapted to perform an image similarity search using the feature extraction MLL to extract feature vectors from a search image and from each of the searched images (tiles), respectively.

[0305] The computer system may be, for example, a standard computer system or server including or operably coupled to a database 992. For example, the database may be a relational BDSM that includes hundreds or thousands of full slide images depicting tissue samples from multiple patients. Preferably, the database includes, for each of the images in the database, a respective feature vector extracted from said image in the database by the feature output MLL 950. Preferably, the computation of the feature vector for each image in the database is performed in a single pre-processing step before any such request is received. However, it is also possible to dynamically compute and extract feature vectors for images in the database in response to a search request. The search may be limited to tiles derived from a particular digital image, for example, to identify tiles within a single full slide image depicting a tissue pattern similar to the tissue pattern depicted in the search image 986. The search image 986 may be, for example, tiles included in a report tile gallery selected by a user.

[0306] The computer system includes a user interface that allows a user 984 to select or provide a particular image or image tile to be used as a search image 986. The trained feature extraction MLL 950 is adapted to extract a feature vector 988 ("search feature vector") from the input image. The search engine 990 receives the search feature vector 988 from the feature output MLL 950 and performs a vector-based similarity search in the image database. The similarity search involves comparing the search feature vector with each of the feature vectors of the images in the database to calculate a similarity score as a function of the two compared feature vectors. The similarity score indicates the degree of similarity of the search feature vector with the feature vectors of the images in the database, and thus the similarity of the tissue patterns depicted in the two compared images. The search engine 990 is adapted to output search results 994 back to the user. The search results can be, for example, one or more images in the database for which the highest similarity score was calculated.

[0307] For example, if the search image 986 is an image tile known to depict breast cancer tissue, the system 980 can be used to identify multiple other tiles (or entire slide images containing such tiles) that depict similar breast cancer tissue patterns.

[0308] FIG. 13 shows two tile matrices, each with three columns and each with six tile pairs. The first (top) matrix shows a first set of tile pairs (A) that are close to each other and have been automatically assigned the label "similar" tile pairs. The second (bottom) matrix shows a second set of tile pairs (B) that are far from each other and have been automatically assigned the label "dissimilar" tile pairs. In some cases, tiles labeled "similar" do not look similar, and tiles labeled "dissimilar" do look similar. This noise is caused by the fact that at the boundary where two different tissue patterns meet, two nearby tiles may depict different tissue patterns, and by the fact that even distant tissue regions may depict the same tissue pattern. This is an expected and inherent noise in the dataset generation process.

[0309] Applicants have observed that, despite this noise, a feature extraction MLL trained on an automatically labeled dataset is able to accurately identify and extract features that allow for clear distinction between similar and dissimilar tile pairs. Applicants hypothesize that the observed robustness of the trained MLL to this noise is due to the fact that region boundaries typically have smaller areas than region non-boundary areas.

[0310] According to an embodiment, the quality of the automatically generated training dataset is such that in a first step, the similarity of tile pairs is assessed using a previously trained similarity network or an ImageNet pre-trained network, followed by generating similarity labels in a second step based on the spatial proximity of the tiles as described herein for embodiments of the invention, followed by correcting the pair labels where strong deviations are observed between the similarity of two tiles determined, one in the first step and the other in the second step.

[0311] FIG. 14 shows feature vectors based on similarity search results extracted by a feature extraction MLL trained on proximity-based similarity labels. The five tumor query tiles are called A, B, C, D, and E. The query tiles were used in an image retrieval task to identify and retrieve five tiles ranked by low to high distances, other than the query slide (A1 to A5, B1 to B5, C1 to C5, D1 to D5, E1 to E5), using feature vectors extracted by a feature extraction MLL trained on data automatically labeled with proximity-based labels. The target class (e.g., tumor) contains only 3% of the searched tiles. Although some of the retrieved tiles look very different from the query tiles (e.g., C3 and C), all of the retrieved tiles, except A4, are verified by an expert pathologist to contain tumor cells (i.e., correct class retrieval).

[0312] Figure 15 shows Approach I for classifying an input image 212. The image 212 is divided into a number of tiles 216, from which a respective feature vector 220 is extracted as described above. The feature vectors are sequentially input to a machine learning model 999 in a MIL program 226. The model is configured to calculate and output a prediction value h 998 and a certainty value 221 for each feature vector input to the model.

[0313] A pooling function 997 is then used to calculate an aggregate prediction 997 as a function of both the prediction 998 and the certainty value 221. For example, the pooling function can aggregate the weighted prediction 228, i.e., the prediction weighted by the certainty value calculated by the MIL model 999 for the tile for which the prediction is calculated. Alternatively, the certainty value can be evaluated by the pooling function to identify a particular prediction, e.g., the prediction for the tile with the highest certainty value among all tiles of the image. The aggregate prediction is then a normalized number, typically between 0 and 1, and is used to classify the image 212. For example, if the aggregate prediction 887 is less than 0,75, the MIL program can classify the image as a member of the "negative class", otherwise as a member of the "positive class".

[0314] FIG. 16 illustrates an alternative approach II for classifying an image 212. In this approach, a pooling function 996 is applied to the feature vectors extracted from the tiles 216 of the image, rather than the predicted values. This pooling function takes into account the certainty values ​​221 calculated for the tiles. The pooling function calculates an aggregate feature vector, also called a "global feature vector" 995. This global feature vector conveys information aggregated from the features and certainty values ​​of the tiles of the image 212. After calculating the global feature vector 995, the vector is input to the machine learning model 999 of the MIL program in the same way as the "regular" feature vectors fv1, fv2, ..., fv2 of the tiles of the image. The MIL program calculates an aggregate predicted value 997 from the global feature vector in the same way that it calculated any of the predicted values ​​998 from each tile-specific feature value, but because the global vector contains information derived from all tiles, the predicted value calculated from the global value is an aggregate predicted value. Finally, the aggregate predicted value is used to classify the image 212. [Explanation of symbols]

[0315] 100 ways 102-116 Steps 200 Image Analysis System 202 Processor 204 Display 206 Image Tile Gallery 207 Classified Images 208 Whole Slide Heat Map 210 Storage medium 212 Digital Images 214 Division Module 216 Bags of Labeled Tiles 218 Feature Extraction Module 220 feature vectors 221 Certainty Value 222 Attention Machine Learning Logic Program 224 Weighting of Feature Vectors 226 Multi-instance Learning Program Numerical relevance score for 228 tiles 229 Classification results 230 GUI Generation Module 232 GUI 300 Table 1 400 Table 2 502 Standard Neural Net 504 Neural network after dropout 600 Feature Extraction MLL Network Architecture 602 image tiles used as input 603 Input Layer 604 Multiple Layers 606 Bottleneck Layer 700 GUI with report tile gallery 702 a first subset of similar tiles of a first texture pattern 704 A second subset of similar tiles representing a second texture pattern. 706 a third subset of similar tiles representing a third organizational pattern 708 A fourth subset of similar tiles representing a fourth organizational pattern 710 A set of selectable GUI elements 712 All Slide Images 714 All Slide Images 716 total slide images 718 All Slide Images 722 Relevance Heatmap 724 Relevance Heatmap 726 Relevance Heatmap 728 Relevance Heatmap 800 GUI containing similarity search tile gallery 802 a first subset of similar tiles of a first texture pattern 804 A second subset of similar tiles representing a second texture pattern. 806 a third subset of similar tiles representing a third organizational pattern 808 A fourth subset of similar tiles representing a fourth organizational pattern 812 All Slide Images 814 All Slide Images 816 total slide images 818 All Slide Images 822 Similarity Heatmap 824 Similarity Heatmap 826 Similarity Heatmap 828 Similarity Heatmap 830 Query Tiles 900 Digital tissue images sliced ​​into multiple tiles 902 Tile T1 904 Tile T2 906 Tile T3 908 First spatial proximity threshold (2D) 910 Second spatial proximity threshold (2D) 912 Pairs of tiles labeled "similar" 913 Pairs of tiles labeled "similar" 914 Tile pairs labeled "dissimilar" 915 Tile pairs labeled "dissimilar" 916 Training Data 932 Digital tissue image aligned to image 900 934 Digital tissue image aligned to image 932 936 First spatial proximity threshold (3D) 938 Second spatial proximity threshold (3D) 940 Tile T4 942 Tile T5 1000 Siamese Network 1002 Subnetwork 1003 Subnetwork 1004 1st input tile 1005 Input layer of the first network N1 1006 Hidden Layer 1007 Proximity-based ("measured") similarity labels 1008 Hidden layer for computing the feature vector of the first input tile 1010 Feature vector extracted from the first input tile 904 1014 Second input tile 1015 Input layer of the second network N2 1016 Hidden Layers 1018 Hidden layer for computing the feature vector of the second input tile 1020 Feature vector extracted from the second input tile 914 1022 input tile pairs 1024 Output layer joining networks N1 and N2 1026 predicted similarity labels 1028 individual data records in the training dataset 950 Feature Extraction MLL 952 distinct input images / tile 954 feature vectors 980 Computer Systems 982 Processor 984 users 986 distinct input images / tile 988 Search Feature Vector 990 Feature Vector Based Search Engine 992 Database containing multiple images or tiles 994 Similarity Search Results Returned 995 Global Feature Vectors 996 Certainty value based pooling function 997 Aggregate forecast value 998 predicted value 999 Prediction Model

Claims

1. A method (100) for classifying tissue images (212, 712-718, 822-828), comprising: - receiving (102) by an image analysis system (200) a plurality of digital images (212, 712-718, 822-828), each of said digital images representing a tissue sample from a patient; - dividing (104) each received image into a set of image tiles (216, 902-906, 940, 942) by said image analysis system; - calculating (106) for each of said tiles a feature vector (220) comprising image features selectively extracted from said tile by said image analysis system; - providing (108) a multi-instance learning (MIL) program (226), the model being a neural network, configured to classify any input image as a member of one of at least two different classes based on the feature vectors extracted from all tiles of the input image; - calculating (110), for each of said tiles, a certainty value (221) indicating the certainty of said model regarding the contribution of said tile's feature vectors to the classification of the image from which said tile was derived, using a dropout technique during testing of said neural network model; - for each image, using a certainty value based pooling function (996) for aggregating (111) the feature vectors extracted from the image by the MIL program (226) into a global feature vector (995) as a function of the certainty values ​​(221) of the tiles of the image, and calculating (112) an aggregate prediction value (997) by inputting the global feature vector into a prediction model (999) of the MIL program (226); - classifying (116) each of said images as a member of one of said at least two different classes based on said aggregate prediction value by said MIL program; The method comprising:

2. - outputting the classification results to the user via a GUI; and / or - outputting said classification results to other application programs; The method of claim 1 further comprising:

3. said MIL program (226) is a binary MIL program, said at least two classes consist of a first class, called the "positive class", and a second class, called the "negative class", - one of the images is classified into the "positive class" if the MIL program predicts that, for at least one of the tiles of the image, the feature vector of that tile contains evidence of the "positive class"; - Any one of the images is classified into the "negative class" if the MIL program predicts that for all tiles of that image, their respective feature vectors contain no evidence of the "positive class"; The method according to claim 1 or 2.

4. the certainty value-based pooling function is used during testing; Providing the MIL program (108) - extracting feature vectors from a set of training tiles generated from a set of training images; - training the MIL program on the feature vectors, thereby using the same certainty value-based pooling function during training as was used during testing, or using a different certainty value-based pooling function during training than the certainty value-based pooling function used during testing; Including, The certainty value-based pooling function used during training is a certainty value-based max pooling function or a certainty value-based average pooling function, and the certainty value-based pooling function used during testing is a certainty value-based max pooling function. The method according to any one of claims 1 to 3.

5. The certainty value-based pooling function is a certainty value-based maximum pooling function, and using the certainty value-based pooling function includes, for each of the images, a sub-method of: max ) as the global feature vector (995). The method according to any one of claims 1 to 4, comprising:

6. the certainty value based pooling function is a certainty value based average pooling function; using the certainty value based pooling function comprises: for each of the images, weighting the feature vector (fv) of each tile by the certainty value (c) calculated for that tile, thereby obtaining a weighted feature vector (wfv), and calculating the global feature vector as an average of all the weighted feature vectors (wfv) of the images. The method according to any one of claims 1 to 5.

7. The method according to any one of claims 1 to 6, wherein the certainty value is calculated as a Monte Carlo Dropout (MC Dropout).

8. The method of claim 7 , wherein the dropout technique and / or the certainty value based pooling function are used at test time rather than at training time of the model.

9. the neural network includes one or more deactivated dropout layers, the deactivated dropout layers being dropout layers that are activated during training and deactivated during testing, and the method includes reactivating the one or more dropout layers during testing; or the neural network at training time does not include any dropout layers, and the method includes adding one or more dropout layers to the neural network at test time; Calculating the certainty value for any one of the tiles at test time comprises: - for each of the tiles, a prediction (h d ) multiple times, d each time a different subset of nodes of the network is dropped by the one or more reactivated or added dropout layers; - for each of the tiles, the plurality of predicted values ​​(h d calculating the certainty value for the tile as a function of variability in (c), the greater the variability, the smaller the certainty value (c); The method of claim 7 or 8, further comprising:

10. The received digital image is - digital images of tissue samples, the pixel intensity values ​​of which correlate with the amount of non-biomarker specific stain, in particular hematoxylin stain or H&E stain, and / or - digital images of tissue samples, the pixel intensity values ​​of which correlate with the amount of a biomarker-specific dye adapted to selectively stain the biomarker contained in said tissue sample; The method according to any one of claims 1 to 9, comprising:

11. providing the MIL program includes training a model of the MIL program; The training, providing a set of digital training images of tissue samples, each digital training image being assigned a class label indicating one of at least two classes; - dividing each training image into training image tiles, each training tile being assigned the same class label as the digital training image from which it was derived; - calculating, for each of said tiles, a training feature vector comprising image features selectively extracted from said tiles by said image analysis system; and / or - iteratively adapting the model of the MIL program such that an error of a loss function that indicates the difference between the predicted class labels of the training tiles and the class labels actually assigned to the training tiles is minimized, the predicted class labels being calculated by the model based on the feature vectors of the training tiles. The method according to any one of claims 1 to 10, comprising:

12. For each of said received digital images, - weighting the prediction value (h) of each tile by the certainty value (c) calculated for that tile, thereby obtaining a weighted prediction value (wh); - identifying one of the tiles of the image for which the highest weighted prediction value (wh) is calculated by the MIL program; - calculating, for each other tile of the image, a relevance indicator by comparing the weighted predicted value (wh) of said other tile with said highest weighted predicted value, said relevance indicator being a numerical value that is negatively correlated with the difference between the compared weighted predicted values; - calculating a relevance heatmap (722-728, 822-828) for said images (712-718, 812-828) as a function of said relevance indicators, where pixel colours and / or pixel intensities of said relevance heatmap indicate the relevance indicators calculated for said tiles in said images; - displaying said relevance heatmap on a GUI; The method of any one of claims 1 to 11, further comprising:

13. - displaying the received images on a screen GUI, the images being grouped in said GUI into said at least two different classes according to the results of said classification; The method of any one of claims 1 to 12, further comprising:

14. Each of the at least two classes comprises: - Patients who are sensitive to certain drugs, - Patients who developed metastases or micrometastases, - cancer patients who have a pathological complete response (pCR) to a particular treatment; - cancer patient tissues exhibiting a particular morphological or microsatellite status, - Patients who have developed adverse reactions to certain medications, - Patients with a particular genetic signature, and / or - Patients with a particular RNA expression profile The method according to any one of claims 1 to 13, wherein the compound is selected from the group comprising:

15. The method of any one of claims 1 to 14, wherein the certainty value based pooling function (996) is a certainty value based max pooling, average pooling, or attention pooling function.

16. - Providing a trained attention MLL (222) that has learned which tile-derived feature vectors are most relevant for predicting the class membership of a tile; - calculating an attention weight (aw) for each of the tiles as a function of the feature vector of the respective tile by the attention MLL, the attention weight being a measure of the relevance of the feature values ​​of that tile with respect to the membership of that tile to a class; - multiplying the attention weight (aw) of the tile by the feature vector value of the tile to obtain an attention-based feature vector having an attention-weighted feature value of the tile; and using the attention-based feature vector as the feature vector input to the MIL program for calculating a slide prediction value (h), the certainty value (c), and / or the weighted prediction value (wh), whereby the prediction value (h), the certainty value (c), and / or the weighted prediction value (wh) are calculated as attention-based prediction value, attention-based certainty value, and / or attention-based weighted prediction value, or - multiplying the attention weight (aw) of the tile by the prediction value (h), the certainty value (c) of the tile or by the weighted prediction value (wh) of the tile calculated by the MIL to obtain the attention-based prediction value, the attention-based certainty value and / or the attention-based weighted prediction value. The method of any one of claims 1 to 15, further comprising:

17. 1. An image analysis system for classifying tissue images, comprising: at least one processor (202); - a volatile or non-volatile storage medium (210) containing digital images each representing a tissue sample of a patient; an image segmentation module (214) executable by said at least one processor and configured to segment each of said images into a set of image tiles (216); a feature extraction module (218) executable by said at least one processor and configured to calculate, for each of said tiles, a feature vector (220) comprising image features selectively extracted from said tiles; a multi-instance learning (MIL) program (226) executable by said at least one processor and configured to use a model, said model being a neural network, for classifying any input image as a member of one of at least two different classes based on said feature vectors extracted from all tiles of the input image; wherein the MIL program comprises: - calculating (110) for each of said tiles, using a dropout technique during testing of said neural network model, a certainty value indicating the certainty of said model regarding the contribution of said tile's feature vectors to the classification of the image from which said tile was derived; - for each image, using a certainty value based pooling function (996) for aggregating (111) the feature vectors extracted from said image by said MIL program (226) into a global feature vector (995) as a function of the certainty values ​​(221) of the tiles of said image, and calculating (112) an aggregate prediction value (997) from said global feature vector; - classifying (116) each of said images as a member of one of said at least two different classes based on said aggregate prediction value by said MIL program; The image analysis system further configured for:

Citation Information

Patent Citations

  • System and method for quantifying uncertainty in reasoning about 2d and 3D spatial features with a computer machine learning architecture

    US20190122073A1