Active Learning System for Digital Pathology

Abstract

A machine learning model configured to process images using one or more parameters to generate labels is accessed. The machine learning model is executed to convert at least a portion of each of the at least one digital pathology image into a plurality of predicted labels and generate a confidence metric for each of the plurality of predicted labels. An interface is utilized that depicts at least a portion of the at least one digital pathology image and differentially represents the predicted labels based on the corresponding confidence metric. In response to utilization of the interface, a label input is received that confirms, rejects, or replaces at least one of the plurality of predicted labels. One or more parameters of the machine learning model are updated based on the label input.

Description

[Technical field]

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS This application claims the benefit of and priority to U.S. Provisional Patent Application No. 63 / 065,404, filed August 13, 2020, which is incorporated by reference herein in its entirety for all purposes.

[0002] Field The present disclosure relates to digital pathology, and in particular to techniques for using active learning to efficiently train machine learning models to automatically detect, characterize and / or classify some or all of digital pathology images. [Background technology]

[0003] background Digital pathology involves scanning a slide (e.g., a histopathology or cytopathology glass slide) into a digital image. The tissues and / or cells in the digital image can then be examined using digital pathology image analysis and / or interpreted by a pathologist for a variety of reasons, including diagnosing disease, evaluating response to treatment, and developing drugs to combat disease. To examine the tissues and / or cells in the digital image (which is substantially transparent), the pathology slide can be prepared using a variety of staining assays (e.g., immunostaining) that selectively bind to specific tissue and / or cellular components.

[0004] Training a machine learning model to analyze digital pathology images requires a large set of ground truth, manually labeled images. For example, manual labeling can include identifying the location (e.g., point location and / or boundary) of each cell of a given type (e.g., tumor cells) within a particular region in an image. Generating these labeled images can be tedious and time-consuming to collect. Furthermore, obtaining training images can be difficult due to privacy concerns.

[0005] Currently, to collect ground truth to prepare a training dataset, a selection of images or a selection of parts of images to be labeled is randomly selected from a pool of data (e.g., available images). However, randomly choosing images or parts or images to be labeled is not an efficient approach. The images or parts of images that are randomly selected samples may not be the most informative in training a machine learning model and may lack depictions of interest. Thus, labeling randomly selected images or randomly selected parts of images may waste resources (e.g., time and computational resources of pathologists involved in model training) without adding any value to the training process. Summary of the Invention

[0006] overview BRIEF DESCRIPTION OF THE DRAWINGS Aspects and features of various embodiments will become more apparent from the following detailed description of the invention, taken in conjunction with the accompanying drawings, in which: FIG.

[0007] In some embodiments, a method for active learning in a digital pathology system is provided. A machine learning model configured to process images using one or more parameters to generate labels is accessed. The machine learning model is executed to convert at least a portion of each of the at least one digital pathology image into a plurality of predicted labels and generate a confidence metric for each of the plurality of predicted labels. An interface is utilized that depicts at least a portion of the at least one digital pathology image and differentially represents the predicted labels based on the corresponding confidence metric. In response to utilizing the interface, a label input is received that confirms, rejects, or replaces at least one of the plurality of predicted labels. One or more parameters of the machine learning model are updated based on the label input.

[0008] The method may further include one or more additional actions.

[0009] The one or more additional operations may include generating a confidence map of the digital pathology image, where a value of each portion of the digital pathology image corresponds to a statistical value of one or more confidence metrics corresponding to a point location or region within the portion, and utilizing the interface includes utilizing a representation of the confidence map.

[0010] The interface may be configured to receive inputs for representing statistics regarding a region confidence metric at one magnification level, for representing the confidence metric at another magnification level, and for receiving inputs for changing the magnification level.

[0011] The one or more additional operations may include receiving a new digital pathology image, converting the new digital pathology image into one or more new labels that characterize the new digital pathology image or a portion of the new digital pathology image, and outputting the one or more new labels.

[0012] The one or more additional operations may include pre-processing the initial version of the digital pathology image to detect a location of each feature of interest, and transforming at least a portion of the digital pathology image includes generating a label corresponding to each detected location of interest within at least a portion of the digital pathology image.

[0013] The one or more additional operations may include, for each of the plurality of predicted labels, selecting a size, color, or shape of a marker based on a confidence metric of the predicted label, and the interface using a marker to represent each predicted label of the plurality of predicted labels using a marker having the selected size, color, or shape, thereby differentially representing the predicted labels based on the corresponding confidence metric.

[0014] The machine learning model can include a deep neural network.

[0015] In some embodiments, a method is provided that includes determining, by a user, a diagnosis of a subject based on results generated by a machine learning model trained using some or all of one or more techniques disclosed herein, and potentially selecting, recommending and / or administering a particular treatment to the subject based on the diagnosis.

[0016] In some embodiments, a method is provided that includes determining, by a user, a treatment to select, recommend and / or administer to a subject based on results generated by a machine learning model trained using some or all of one or more techniques disclosed herein.

[0017] In some embodiments, a method is provided that includes determining, by a user, whether a subject is eligible to participate in a clinical trial or assign the subject to a particular cohort in a clinical trial based on results generated by a machine learning model trained using some or all of one or more of the techniques disclosed herein.

[0018] In some embodiments, a system is provided that includes one or more data processors and a non-transitory computer-readable storage medium containing instructions that, when executed on the one or more data processors, cause the one or more data processors to perform some or all of one or more of the methods disclosed herein.

[0019] In some embodiments, a computer program product is provided that is tangibly embodied in a non-transitory machine-readable storage medium and includes instructions configured to cause one or more data processors to perform some or all of one or more of the methods disclosed herein.

[0020] The terms and expressions employed are used as terms of description and not of limitation, and in the use of such terms and expressions there is no intention to exclude any equivalents of the features shown and described or portions thereof, but it is recognized that various modifications are possible within the scope of the invention as claimed. Thus, although the invention as claimed has been specifically disclosed by embodiments and optional features, it will be understood that modifications and variations of the concepts disclosed herein may be resorted to by those skilled in the art, and such modifications and variations are considered to be within the scope of the invention as defined by the appended claims.

[0021] The patent or application file contains at least one drawing executed in color. Copies of this patent or patent application publication with color drawing(s) will be provided by the Office upon request and payment of the necessary fee.

[0022] BRIEF DESCRIPTION OF THE DRAWINGS Aspects and features of various embodiments will become more apparent from the following detailed description of the invention, taken in conjunction with the accompanying drawings, in which: FIG. [Brief description of the drawings]

[0023] [Figure 1] 1 shows an example of a tissue staining process. [Diagram 2] 1 illustrates a process flow (performed using one or more computing systems) for processing digital pathology images using machine learning models. [Diagram 3] 1 illustrates an active learning workflow in accordance with some aspects of the present disclosure. [Figure 4] 1 illustrates an exemplary workflow for using some or all of the digital pathology process flow. [Diagram 5] A network 500 is disclosed having exemplary systems and applications that may be used in this regard. [Figure 6]6A-6D show example images of part or all of an entire slide image to facilitate receiving labels for training a machine learning model. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0024] Detailed Description Although specific embodiments are described, these embodiments are presented by way of example only and are not intended to limit the scope of protection. The devices, methods, and systems described herein can be embodied in various other forms. Furthermore, various omissions, substitutions, and modifications of the form of the exemplary methods and systems described herein can be made without departing from the scope of protection.

[0025] I. Overview In various embodiments, an active learning system for digital pathology is provided that dynamically selects specific elements in a dataset to be labeled to train a machine learning model (e.g., a deep learning or traditional machine learning model). The specific elements can include a given slide, a given patch in an image, a given depiction of a cell, etc. The selected specific elements can include an incomplete subset of elements in the dataset (e.g., such that only a portion of the cell depictions are selected for labeling). The selection of the specific elements can help efficiently train the machine learning model to generate accurate predictions. For example, the selection of the specific elements can enable the model to achieve a threshold accuracy with only a portion (e.g., less than 75%, less than 50%, or less than 25%) of the labeled data that would be required to achieve the threshold accuracy if the data elements were randomly selected for labeling. The use of the active learning system allows the model to be trained using less labeled data compared to using randomly selected training data. Furthermore, the selection of the specific elements to be labeled and the training of the machine learning model can be performed iteratively, and weaknesses in the model are repeatedly detected and improved.

[0026] The digital pathology active learning system can interact with and / or be used by one or more pathologists or one or more image scientists. For example, at each of one or more time points, the active learning system can transmit, present, or identify specific elements selected for labeling to a pathologist's device. The pathologist can then provide input including labels that can be used to train a machine learning model. The trained machine learning model can then be used to process input data (e.g., provided or identified by an entity that may be a pathologist or a different entity) and generate label predictions that are output to a device associated with the input data.

[0027] In various embodiments, the active learning system includes high-resolution image storage and retrieval, data visualization, interactive data correction, and image analysis for algorithm training and inference. An initial labeled dataset is used to initially train a machine learning model (either a model with uninitialized parameters or a model with parameters initialized using a pre-training or random initialization process). The initial labeled dataset can include labels of randomly or pseudo-randomly selected elements. The model is then used to generate a set of predictions corresponding to other elements.

[0028] Each prediction may include a predicted label and a corresponding confidence for the label. An interface (e.g., using a scaled visual representation or text) may be utilized for a user (e.g., a pathologist or image scientist) to identify the selected elements to be labeled (or for label review) and / or to represent the degree to which labeling (or label review) of various elements is predicted to facilitate training relative to other elements. For example, a heat map corresponding to the entire slide image may be presented, with colors representing the confidence of predictions associated with various regions within the depicted slice. The interface may allow a user to zoom in on a given field of view to enlarge a particular region (e.g., to confirm or reject a label, to positively identify a label, etc.) and review the predicted label. These user inputs may be used to retrain and / or continue training the model (e.g., to update one or more parameters of the model). In some cases, the interface dynamically updates itself so that user-identified or user-verified labels are associated with high (or maximum confidence) and / or elements associated with the lowest confidence metrics (e.g., relative to others in the field of view, in the whole slide image, or in the dataset) are prominently identified (e.g., via a given color, marker type, line type, line width, etc.).

[0029] In some cases, the model is periodically further trained based on batches of user input (e.g., identifying element labels or providing feedback on predicted labels). In this case, an interface may be periodically updated or periodically provided to identify predicted labels generated based on output from the latest version of the model.

[0030] The iterations between generating predicted labels, providing user feedback corresponding to the most uncertain predictions, and retraining the model can continue for a predetermined period of time, a predetermined number of iterations, or until the model reaches a desired performance (at which point the model can be deployed). The iterations can be continued at later times to further fine-tune the classifier.

[0031] The high performance image server can support one or more applications for building, training, and / or using models. Results from model training and inference can be generated and stored at scale in a data store. An API can enable interaction between the model controller (which trains and / or executes the models) and the data store. An integration platform can be used to integrate all of these components, allowing users to train models within a web browser.

[0032] II. Definition As used herein, when an action is "based on" something, this means that the action is at least somewhat, or at least partially, based on something.

[0033] As used herein, the terms "substantially," "approximately," and "about" are defined as being largely, but not necessarily entirely, specified (and including) that which is specified, as would be understood by one of ordinary skill in the art. In any disclosed embodiment, the terms "substantially," "approximately," or "about" can be substituted with "within [a percentage]" of that which is specified, where the percentage includes 0.1, 2, 5, and 20%.

[0034] As used herein, the terms "sample", "biological sample" or "tissue sample" refer to any sample containing biomolecules (e.g., proteins, peptides, nucleic acids, lipids, carbohydrates, or combinations thereof) obtained from any biological material, including viruses, and / or organisms. Other examples of organisms include mammals (such as veterinary animals, such as humans, cats, dogs, horses, cows, and pigs, and laboratory animals, such as mice, rats, primates, etc.), insects, annelids, arachnids, marsupials, reptiles, amphibians, bacteria, and fungi, etc. Biological samples include tissue samples (such as tissue sections or needle biopsies of tissues), cell samples (such as cytological smears, such as Pap smears or blood smears, or samples of cells obtained by microdissection), or cell fractions, fragments, or organelles (obtained by lysing cells and separating their components, such as by centrifugation). Other examples of biological samples include blood, serum, urine, semen, feces, cerebrospinal fluid, interstitial fluid, mucous membranes, tears, sweat, pus, biopsy tissue (e.g., obtained by surgical or needle biopsy), nipple aspirate, earwax, milk, vaginal fluid, saliva, swabs (such as cheek swabs), or any material that contains biomolecules derived from an initial biological sample. In certain embodiments, the term "biological sample" as used herein refers to a sample prepared from a tumor or portion thereof obtained from a subject (such as a homogenized or liquefied sample).

[0035] As used herein, the term "biological material or structure" refers to a naturally occurring material or structure that contains all or part of a biological structure (e.g., a cell nucleus, cell membrane, cytoplasm, chromosome, DNA, cell, cell mass, etc.).

[0036] As used herein, "digital pathology image" refers to a digital image of a stained sample.

[0037] As used herein, the term "artifact" refers to artificial structures or tissue changes in a prepared microscope slide as a result of exogenous factors. Artifacts can occur, for example, during surgical removal, fixation, tissue processing, embedding, sectioning, staining, and mounting procedures. Artifacts in histopathology can include, for example, prefixation artifacts, fixation artifacts, bone tissue-related artifacts, tissue processing artifacts, sectioning-related artifacts, floating and mounting-related artifacts, staining artifacts, mounting artifacts, biological artifacts, imaging artifacts, and the like. For example, prefixation artifacts can include injection artifacts, squeeze artifacts (e.g., tissue tears or tissue folds), sulfide artifacts, starch artifacts, autolysis artifacts, and the like. Fixation artifacts can include, for example, formalin pigments, mercury pigments, ice crystal artifacts, freezing artifacts, streaming artifacts, and the like. Floating and mounting-related artifacts can include, for example, creases and wrinkles in the cross section, contamination, air bubbles, and the like. Staining artifacts can include, for example, residual wax, artifacts associated with the addition of acetic acid to eosin, artifacts due to the mordanting of hematoxylin, artifacts due to the fluorescent sheen of hematoxylin, etc. Mounting artifacts can include, for example, residual water and air bubbles. III. Generation of digital pathology images

[0038] Histological staining is widely used to highlight features of interest and enhance the contrast of sectioned tissues or cells of biological samples. For example, staining can be used to mark specific types of cells and / or flag specific types of nucleic acids and / or proteins to aid in microscopy. The stained sample can then be evaluated to determine or estimate the amount of the feature of interest in the sample (which can include, for example, counts, density or expression levels) and / or one or more characteristics of the feature of interest (e.g., the location of the features of interest relative to each other or relative to other features, shape characteristics, etc.). The process of histological staining can include several stages, such as fixing, processing, embedding, sectioning, staining and imaging.

[0039] In some embodiments, immunohistochemical staining of tissue sections is a type of histological staining used to identify the presence of a particular protein in a biological sample. For example, the expression level of a particular protein (e.g., an antigen) is determined by: (a) performing immunohistochemical analysis of a sample with a particular antibody type; (b) determining the presence and / or expression level of the protein in the sample. In some embodiments, immunohistochemical staining intensity is determined relative to a reference sample (e.g., a control cell line staining sample, a tissue sample from a non-cancerous subject, a reference sample known to have a predetermined level of protein expression).

[0040] FIG. 1 shows an example of a histological staining process 100. Stage 110 of the histological staining process 100 includes sample fixation, which can be used to preserve the sample and slow the degradation of the sample. In histology, fixation generally refers to an irreversible process that uses chemicals to preserve the chemical composition, preserve the native sample structure, and maintain cellular structures from degradation. Fixation can also harden cells or tissues for sectioning. Fixatives can enhance sample and cell preservation using crosslinking proteins. Fixatives can bind and crosslink some proteins and denature others by dehydration, which can harden the tissue and inactivate enzymes that might otherwise degrade the sample. Fixatives can also kill bacteria.

[0041] Fixatives can be administered, for example, by perfusion and immersion of the prepared sample. Various fixatives can be used, including methanol, buine fixatives and / or formaldehyde fixatives, such as neutral buffered formalin (NBF) or paraffin-formalin (paraformaldehyde-PFA). If the sample is a liquid sample (e.g., a blood sample), the sample may be smeared onto a slide and dried before fixation.

[0042] Although the fixation process can help preserve the structure of samples and cells for the purposes of histological examination, fixation can result in the concealment of tissue antigens, thereby reducing antigen detection. Thus, fixation is generally considered a limiting factor in immunohistochemistry, since formalin can crosslink antigens and mask epitopes. In some cases, additional processes are performed to reverse the effects of crosslinking, including treating fixed samples with citraconic anhydride (a reversible protein crosslinker) and heating.

[0043] Stage 120 of the histological staining process 100 includes sample processing and embedding. Sample processing can include infiltrating a fixed sample (e.g., a fixed tissue sample) with a suitable histological wax, such as paraffin wax. The histological wax may be insoluble in water or alcohol, but may be soluble in a paraffin solvent, such as xylene. Thus, it may be necessary for the water in the tissue to be replaced by xylene. To do so, the sample can be first dehydrated by gradually replacing the water in the sample with alcohol, which can be achieved by passing the tissue through increasing concentrations of ethyl alcohol (e.g., from 0 to about 100%). After the water has been replaced by the alcohol, the alcohol may be replaced by xylene, which is miscible with the alcohol. Embedding can include embedding the sample in warm paraffin wax. Because paraffin wax can be soluble in xylene, the molten wax can be filled by xylene to fill the spaces previously filled by water. The wax-filled sample can be cooled to form a hardened block that can be clamped to a microtome for sectioning. In some cases, deviations from the above exemplary procedure result in penetration of the paraffin wax resulting in inhibition of penetration of antibodies, chemicals, or other fixatives.

[0044] Stage 130 of the histological staining process 100 involves sectioning the sample. Sectioning is the process of cutting thin slices of a sample from an embedded block for the purpose of mounting the sample (e.g., an embedded and fixed tissue sample) on a microscope slide for examination. Sectioning can be performed using a microtome. In some cases, the tissue can be quickly frozen in dry ice or isopentane and then cut by a cold knife in a refrigerated cabinet (e.g., a cryostat). Other types of coolants, such as liquid nitrogen, can be used to freeze the tissue. Sections for use by light microscopy are generally on the order of 4-10 μm thick. In some cases, the sections can be embedded in epoxy or acrylic resin, which can allow thinner sections (e.g., <2 μm) to be cut. The sections can be placed on glass slides.

[0045] Stage 140 of the histological staining process 100 involves staining (of tissue samples or sections of fixed liquid samples). The purpose of staining is to identify different sample components by color reaction. Most cells are colorless and transparent. Therefore, it may be necessary to stain histological sections to visualize the cells. The staining process generally involves adding a dye or staining agent to the sample to confirm or quantify the presence of a particular compound, structure, molecule, or feature (e.g., subcellular feature). For example, staining can help to identify or highlight a particular biomarker from a tissue section. In other examples, staining can be used to identify or highlight biological tissues (e.g., muscle fibers or connective tissue), cell populations (e.g., different blood cells), or organelles within individual cells.

[0046] Many staining solutions are water-based. Therefore, to stain tissue sections, the wax may need to be dissolved and replaced by water (rehydrated) before applying the staining solution to the sections. For example, sections can be passed sequentially through xylene, decreasing concentrations of ethyl alcohol (from about 100% to 0%), and water. Once stained, sections may be dehydrated again and placed in xylene. Sections can then be mounted on microscope slides in mounting medium dissolved in xylene. To protect the sample sections, a coverslip can be placed on top. Evaporation of the xylene around the edges of the coverslip dries the mounting medium and allows the coverslip to be firmly bonded to the slide.

[0047] Various types of staining protocols can be used to perform staining. For example, an exemplary immunohistochemical staining protocol includes using a hydrophobic barrier line around the sample (e.g., tissue section) to prevent leakage of reagents from the slide during incubation, treating the tissue section with reagents to block endogenous sources of nonspecific staining (e.g., enzymes, free aldehyde groups, immunoglobulins, other unrelated molecules that may mimic specific staining), incubating the sample with a permeabilization buffer to facilitate penetration of antibodies and other staining reagents into the tissue, incubating the tissue section with a primary antibody at a specific temperature (e.g., room temperature, 6-8 °C) for a period of time (e.g., 1-24 hours), rinsing the sample using a wash buffer, incubating the sample (tissue section) with a secondary antibody at another specific temperature (e.g., room temperature) for another period of time, rinsing the sample again using a water buffer, incubating the rinsed sample with a chromogen (e.g., DAB), and washing off the chromogen to stop the reaction. In some cases, a counterstain is then used to distinguish the overall "landscape" of the sample and serve as the primary color reference used to detect tissue targets. Examples of counterstains include hematoxylin (blue to purple stain), methylene blue (blue stain), toluidine blue (nuclear deep blue, polysaccharide pink to red stain), nuclear fast red (also called Kernechtrot dye, stains red), methyl green (green stain), non-nucleogenic stains such as eosin (pink stain), etc. Those skilled in the art will recognize that other immunohistochemical staining techniques can be implemented to perform the staining.

[0048] In another example, an H&E staining protocol can be performed for staining tissue sections. The H&E staining protocol involves applying hematoxylin stain mixed with a metal salt or mordant to the sample. The sample can then be rinsed with a weak acid solution to remove excess stain (differentiation), followed by blueing in weak alkaline water. After application of hematoxylin, the sample can be counterstained with eosin. It will be understood that other H&E staining techniques can be performed.

[0049] In some embodiments, staining can be performed using various types of stains depending on which feature of interest is. For example, DAB can be used on various tissue sections for IHC staining, where DAB results in a brown color that shows the feature of interest in the stained image. In another example, alkaline phosphatase (AP) can be used on skin tissue sections for IHC staining, since the DAB color can be masked by melanin pigment. With regard to the primary staining technique, applicable stains can include, for example, basophilic and acidophilic stains, hematin and hematoxylin, silver nitrate, trichrome stains, etc. Acidic dyes can react with cationic or basic components in tissues or cells, such as proteins and other components in the cytoplasm. Basic dyes can react with anionic or acidic components in tissues or cells, such as nucleic acids. As mentioned above, one example of a staining system is H&E. Eosin can be a negatively charged pink acidic dye, and hematoxylin can be a purple or blue basic dye that contains hematein and aluminum ions. Other examples of stains include Periodic Acid-Schiff (PAS) stains, Masson's Trichrome, Alcian Blue, Van Gieson, Reticulin stains, etc. In some embodiments, different types of stains can be used in combination.

[0050] Stage 150 of the histological staining process 100 includes medical imaging. A microscope (e.g., an electron microscope or an optical microscope) can be used to magnify the stained sample. For example, an optical microscope can have a resolution of less than 1 μm, such as about several hundred nanometers. To observe finer details in the nanometer or sub-nanometer range, an electron microscope can be used. A digital image of the magnified sample can be captured using a digital camera (e.g., integrated with or connected to the microscope). The digital image can include a whole slide image depicting the entire slide and / or the entire stained tissue slice.

[0051] At step 160, images of the stained sections are stored. The images can be stored locally, remotely, and / or on a cloud server. Each image may be stored in association with a subject identifier and a date (e.g., the date the sample was collected and / or the date the image was captured). The images may also be transmitted to another system (e.g., a system associated with a pathologist, or an automated or semi-automated image analysis system).

[0052] It will be understood that modifications to process 100 are contemplated. For example, if the sample is a liquid sample, step 120 (processing and embedding) and / or step 130 (sectioning) may be omitted from the process. IV. Exemplary Process Flow for Digital Pathology Image Conversion

[0053] 2 illustrates a process flow 200 (performed using one or more computing systems) for processing a digital pathology image using a machine learning model. As described further herein, processing the digital pathology image may include training a machine learning model using the digital pathology image or converting some or all of the digital pathology image into one or more results using a trained (or partially trained) version of the machine learning model.

[0054] As shown in FIG. 2, the process flow 200 includes several stages: an image storage stage 205 , a pre-processing stage 210 , a labeling stage 215 , an (active) model training stage 220 , and a result generation stage 225 .

[0055] The image storage stage 205 may include one or more image data stores 260 that are accessed (e.g., by the pre-processing stage 210) to provide a set of digital images 245 of preselected regions from a biological sample slide (e.g., a tissue slide) or from the entire biological sample slide. Each digital image 245 stored in each image data store 260 and accessed in the image storage stage 210 may include a digital pathology image generated according to some or all of the process 100 shown in FIG.

[0056] In the pre-processing stage 210, each, a plurality, or all of the set of digital images 245 are pre-processed using one or more techniques to generate corresponding pre-processed images 250. Pre-processing may include cropping the images. In some cases, pre-processing may further include standardization or rescaling (e.g., normalization) to bring all features to the same scale (e.g., same size scale or same color or saturation scale). In a particular example, the images are resized with a minimum size (width or height) of a given number of pixels (e.g., 2500 pixels) or a maximum size (width or height) of a given number of pixels (e.g., 3000 pixels) to maintain the original aspect ratio.

[0057] In some cases, pre-processing includes locating and / or segmenting each feature of interest in the digital image. Locating the features of interest can include predicting point locations and / or boundaries of the image. For example, localization can include predicting point locations corresponding to each depicted nucleus or predicting closure shapes corresponding to each depicted cell membrane. Feature localization can be performed automatically using one or more automatic detection techniques.

[0058] In some cases, localization is performed semi-automatically or automatically (by feature localization controller 252). For example, intensities across an image (e.g., a preprocessed image) can be normalized or regularized, intensities can be thresholded or filtered, and / or algorithms (e.g., configured to detect objects, lines, and / or shapes) can be applied. Each boundary output and / or point location can then be identified as a feature of interest. In some cases, a metric is associated with each feature of interest (e.g., a metric indicative of the confidence of identifying the feature of interest), and the interest level can be scaled based on the metric.

[0059] In some cases, localization is performed using another image corresponding to the preprocessed image 250, or using a different preprocessed image corresponding to the same base image 245. For example, for a given digital pathology image, a given preprocessing operation can be performed to enhance colors corresponding to dyes absorbed by the nucleus or cell membrane, and the preprocessed image can be used to localize the cells. On the other hand, if each depicted cell is to be classified (e.g., via labeling or using a machine learning model) as a given cell type, a different preprocessing can be used to ensure that other cell organelles are depicted. As another example, consecutive slices of a sample may be stained using different stains, such that some of the slices are stained to facilitate localization of features of interest (e.g., using H&E staining to facilitate cell detection) and adjacent slices are stained to facilitate cell classification by staining for different biomarkers. In this case, localization of features of interest can be mapped, transformed, and / or mapped from the image used to identify the localization to the image used for labeling.

[0060] Some preprocessed images 250 are partially or completely labeled in a labeling stage 215. The labels assigned in the labeling stage 215 are "ground truth" labels identified based on input from a human user (e.g., a pathologist or image scientist). Thus, the labeling stage can include transmitting and / or presenting some or all of one or more preprocessed images to a user-operated device. In some cases, the labeling stage 215 includes utilizing an interface (e.g., using an API) presented on the user-operated device, the interface including an input component for accepting input identifying a label. For example, the input component can include a drop-down menu, a radio button menu, a text box (e.g., configured to receive one or more characters), a pencil tool (e.g., configured to identify point locations, curves, boundaries, etc.), etc.

[0061] In some cases, the interface can identify which and / or how much of a particular label is requested, which can be conveyed via (for example) text instructions and / or visualization. For example, a particular color, size, and / or symbol can represent that a label is requested for a particular representation (e.g., a particular cell or region) in the image relative to other representations. If labels corresponding to multiple representations are requested, the interface can identify each representation simultaneously (such that providing a label for one identified representation triggers the identification of the next representation for labeling) or can identify each representation sequentially.

[0062] In some cases, each image is presented until the user identifies a certain number of labels (e.g., of a certain type). For example, a given full slide image or a given patch of a full slide image can be presented until the user identifies 15 locations, each of which corresponds to a depiction of a macrophage, at which point the interface can present a different full slide image or image of a different patch (e.g., until a threshold number of images or patches have been labeled). Thus, in some cases, the interface is configured to request and / or accept labels of an incomplete subset of features of interest, and the user can determine which of the potentially many depictions are labeled.

[0063] The interface may support zooming in and out of the image and / or moving the image (e.g., to magnify various fields of view or back to full slide scale). In some cases, at low magnifications, portions of the image for which labels are required may be identified using one type of visual identifier (e.g., a color map), and at high magnifications, portions of the image for which labels are required may alternatively or additionally be identified using another type of visual identifier (e.g., a marker symbol, marker size, and / or marker shape). In some cases, characteristics of the visual indicator of the portion of the image for which labels are required change with magnification. For example, the size of the marker at the point location may be correlated with the magnification level.

[0064] The labels 255 identified in the labeling stage 215 may identify (for example) a classification of cells, a binary indication as to whether a given cell is a particular type of cell, a binary indication as to whether the preprocessed image 250 (or a particular region having the preprocessed image 250) contains a particular type of depiction (e.g., necrosis or artifacts), a slide-level or region-specific categorical characteristic of the depiction (e.g., identifying a particular type of artifact), a number (e.g., identifying the amount of a particular type of cell in a region, the amount of depicted artifacts, or the amount of necrotic region), etc. In some cases, the labels 255 include a location. For example, the labels 255 may identify a point location of a nucleus of a particular type of cell or a point location of a particular type of cell. As another example, the labels 250 may include edges or boundaries of depicted tumors, blood vessels, necrotic regions, etc. Depending on the features of interest, a given labeled preprocessed image 250 may be associated with a single label 255 or multiple labels 255. In the latter case, each label 255 may be associated with (for example) an indication as to which location or portion within the preprocessed image 250 it corresponds to.

[0065] In the labeling stage 215, the labels 255 and the corresponding preprocessed images 250 can be used to train a machine learning model 260. The machine learning model 260 can include (for example) a convolutional neural network ("CNN"), a deep neural network, an initial neural network, a residual neural network ("Resnet"), a U-Net, a V-Net, a single shot multi-box detector ("SSD") network, a recurrent neural network ("RNN"), a rectified linear unit ("ReLU"), a long short term memory ("LSTM") model, a gated regression unit ("GRU") model, or the like, or any combination thereof. The machine learning model can include a decision tree model, a random forest model, a support vector machine, and / or a regression model.

[0066] A machine learning model can include and / or be configured with one or more hyperparameters that are fixed (e.g., defined by a programmer). For example, the hyperparameters can define the number of layers in a neural network, the number of nodes in a layer, the learning rate, etc.

[0067] Training the machine learning model 260 may include learning a set of parameters (e.g., one or more coefficients and / or weights) using a loss or objective function (e.g., to minimize a loss or maximize an objective). Each parameter may be a tunable variable such that the parameter's value is adjusted during training. For example, the loss or objective function can be configured to optimize accurate classification of the depicted representation, optimize characterization of a given type of feature (e.g., characterizing shape, size, uniformity, etc.), optimize detection of a given type of feature, and / or optimize accurate localization of a given type of feature.

[0068] Training at any stage may include requesting a human user to label an incomplete subset of the available data. In some cases, during initial training iterations, the subset can be identified using a random or pseudorandom computational algorithm or by requesting the user to select the subset to be labeled.

[0069] For the first training iteration, the machine learning model 260 may (but need not) be initialized with parameter values ​​that were set randomly or pseudo-randomly before the first training. In some cases, the initial parameter values ​​are defined using transfer learning, where the parameters were learned for another task. In some cases, the parameters are defined using transfer learning instead of an initial training phase that involves random, pseudo-random, or fully discriminatory user selection of a subset of the features to be labeled.

[0070] For subsequent iterations, parameters can be set according to parameter definitions from previous iterations, and training performed in the (active) model training stage 220 can be more active in selecting or prioritizing data to be labeled in the labeling stage 215. More specifically, since at least a temporary definition of the parameters of the machine learning model 260 is identified by the first (or subsequent) iteration, the model 260 can be used in the result generation stage 225 to convert other preprocessed images 250 into predicted labels 265. Each predicted label 265 can be associated with the image and / or part of the image used to generate the predicted label 265, as well as a confidence metric 270.

[0071] An interface utilized by a user's device (e.g., to a pathologist) can identify or present portions of images associated with low confidence metrics (e.g., below a predetermined absolute or relative threshold). For example, if the labels correspond to entire images, the active model training can include presenting a predetermined number of images associated with a lowest confidence metric 270 from a given training iteration and / or presenting all images associated with a confidence metric 270 from a given training iteration below a predetermined threshold. As another example, an initial group of images associated with a confidence metric 270 from a given training iteration below a predetermined threshold can be identified, and a subset (having a predetermined number of images) of the initial group of images can be randomly or pseudo-randomly selected and presented.

[0072] The interface may include one or more features described herein to facilitate labeling. For example, the interface may have one or more input components for accepting input identifying a label and / or a zoom function. As another example, the interface may use color (e.g., via a heat map) or other visual indicator to indicate which portion of the image a label or label review is requested for. The interface may present each image or each image portion in association with an indication of the corresponding predicted label 265. For example, the color of a marker identifying a point location of the image portion may identify the value of a binary or categorical predicted label. As another example, a textual heading or overlay may identify the predicted label of the image.

[0073] In some cases, rather than selecting an image or a portion of an image for labeling or label review, the interface is configured to support labeling or label review for all images or all portions of one or more images used to train the model in a given training image. For example, the interface can present visual identifiers for each of many predicted labels 265 simultaneously or sequentially and can accept input of such predicted labels 265 that confirm or reject the predicted labels 265. In this case, the interface may nevertheless highlight predicted labels associated with low or relatively low confidence and / or explicitly request label review of predicted labels associated with low or relatively low confidence. For example, a marker can be overlaid on each portion of the image in the field of view. The color of the marker can represent the predicted label 265, and the size of the marker can be inversely correlated with the corresponding confidence metric 270.

[0074] When a user interacts with the interface to identify a new label for an image part or image, the new label is considered to be the ground truth label 255 for the image part or image. When a user interacts with the interface to confirm a given predicted label for an image part or image, the ground truth label 255 for the image part or image is defined to be the predicted label. When a user interacts with the interface to reject a predicted label, if only two potential labels are available (meaning the labeling is binary), the ground truth label 255 for the image part or image is defined to be the other of the two potential labels.

[0075] The new ground truth labels 255 can then be used to train the machine learning model 260 during the next training iteration. For example, the machine learning model 260 may be initialized with parameter values ​​learned during a previous iteration, and then the model may be trained using the new ground truth labels 255. As another example, the machine learning model 260 can be retrained using labels 255 identified in association with multiple labeling sessions and / or multiple training iterations, but the labels 255 associated with the most recent labeling session can be weighted more highly than other labels.

[0076] The iteration between the generation of predicted labels 265 and the reliability metric 270 (result generation stage 225) provides an interface that facilitates labeling or label review of low reliability predictions, and training the machine learning model 260 with new labels can continue until a stopping condition is met. The training completion condition can be configured to be met (for example) when a predetermined number of iterations are completed, a statistic generated based on the reliability metric 270 (e.g., the average or median reliability metric, or the percentage of reliability metrics above a certain value) exceeds a predetermined reliability threshold, the percentage of reviewed and rejected labels in the most recent label review falls below a predetermined label rejection threshold, and / or when a user device that was involved in the label review closes the label review application. In some cases, a new training iteration can be initiated in response to receiving a corresponding request from a user device.

[0077] The trained machine learning model 260 can then be used (in the result generation stage 225) to process the new preprocessed images 250 to generate new predicted labels 265 and potentially new confidence metrics 270. The new predicted labels (and potentially new confidence metrics 270) can be presented using the same or similar type of interface as used for the initial labeling or label review. In various cases, the interface presenting the new predicted labels may include or lack an input component for rejecting or replacing the predicted labels 265.

[0078] In some cases, analysis controller 280 generates one or more analysis results 285 that are made available to the entity that requested processing of the underlying image. Analysis results 285 may be provided instead of or in addition to the predicted labels 265. Analysis results 285 can be based on the predicted labels. For example, analysis results 285 can identify a number or percentage of depicted cells associated with a particular label.

[0079] It will be appreciated that the process flow 200 is exemplary and that process flows 200 using different stages and / or components are contemplated. For example, in some cases, the network may omit the pre-processing stage 210 such that the images used to train the model and / or the images processed by the model are raw images (e.g., from an image data store). As another example, it will be appreciated that each of the pre-processing stage 210 and the (active) model training stage 220 may include a controller for performing one or more operations described herein. Similarly, the labeling stage 215 is shown in association with the feature location controller 252 and the result generation stage 225 is shown in association with the analysis controller 280, although the controllers associated with each stage may additionally or alternatively facilitate other operations described herein other than feature location and / or generation of analysis results. As yet another example, the depiction of process flow 200 shown in Figure 2 lacks depicted representations of equipment associated with a programmer (e.g., who selected the architecture of the machine learning model 260, which defined how various interfaces work, etc.), equipment associated with a user who provides an initial label or label review (e.g., in the labeling stage 215), and equipment associated with a user who requests model processing of a given image (which may be the same user or a different user than the user who provided the initial label or label review). Despite the lack of depiction of these equipment, process flow 200 may include the use of one, more, or all of the equipment, and may in fact include the use of multiple equipment associated with a corresponding number of users who provide the initial label or label review, and / or a corresponding number of users who request model processing of various images.

[0080] IV.A. Exemplary Training of Models for Labeling Parts of Whole Slide Images 3 illustrates an active learning workflow 300 according to some aspects of the present disclosure. The operations depicted in the active learning workflow 300 can be performed using one or more computing systems (e.g., a model training computing system and / or one or more devices of one or more users providing labels or label reviews). Some or all of the operations performed in the active learning workflow 300 may be performed during the labeling stage 215 and / or the (active) model training stage 220 of the process flow 200 (shown in FIG. 2).

[0081] The active learning workflow includes both an initialization operation 305 and an active learning operation 310. In this illustrative example, a pre-processing operation is performed in block 315 to detect features within the whole slide image (e.g., using a framework such as that represented as VDPSDK, which can support executing a set of commands or scripts in a specific order). Features can include (for example) cells, artifacts, nuclei, etc.

[0082] At block 320, one or more full slide images are opened. Opening a full slide image may include loading a full slide image from a local or remote storage device onto a user device. A full slide image may be opened in response to a user (e.g., a pathologist, an image scientist, or other suitable user) providing a particular type of input within the viewing software (e.g., initiating a session and / or identifying a full slide image). Opening a full slide image may trigger the full slide image to be presented on the user device.

[0083] At block 325, results from an initial algorithm are accessed, the results corresponding to an entire slide image. The initial algorithm can include a machine learning model, which can include a model initialized with randomly or pseudo-randomly selected parameter values. The initial algorithm can include a machine learning model having parameter values ​​defined using transfer learning (i.e., trained in a context corresponding to a different task). The results can include one or more predicted labels.

[0084] At block 330, one or more fields of view (FOV) are selected. Each FOV may identify a portion of the entire slide image. The FOVs may be selected automatically (e.g., using random or pseudo-random techniques or using a predetermined order) or may be selected based on user input (e.g., by clicking on a portion of the presentation of the entire slide image or by selecting a region of the entire slide image). Selecting a FOV may cause the user device to present the FOV (e.g., instead of or in addition to the image of the entire slide).

[0085] At block 335, one, more, or all of the features of interest (identified at block 315) present in the field of view are labeled. Each of the labels can be defined based on input received from a user. Each of these labels can be defined to be a ground truth label 340, which is stored in association with data regarding the corresponding feature at block 345.

[0086] At block 350, a machine learning model and features (e.g., one or more nuclear features, one or more cellular features, one or more morphological features, and / or one or more structural features) to be used for training are selected. For example, the selection may be made based on input received at a user device that identifies a particular type of model from among multiple types of models presented in a list, pull-down menu, or the like. The features may be selected automatically or based on user input. Each of the multiple types of models may differ from one another in terms of model architecture, parameter initialization (e.g., determined based on training in different contexts), etc. The selection may further include defining one or more hyperparameters of the model (e.g., based on input received at the user device). As another example, a model may be selected from among multiple models using random, pseudo-random, or a predefined order.

[0087] At block 355, the selected machine learning model is trained using the ground truth labels 340 and the corresponding portion of the field of view. In some cases, the ground truth labels are split into two subsets, a first subset for training the model (at block 355) and a second subset for validation. One or more validation metrics can be calculated using the second subset of the ground truth labels, and at block 360, it can be determined whether the validation metrics indicate at least a baseline performance quality. If not, the workflow 300 can return to block 350 and another model can be selected.

[0088] If it is determined that the validation metrics indicate at least the baseline performance quality, the workflow 300 transitions to an active learning operation 310 .

[0089] In block 365, the machine learning model with the parameters selected in block 350 and learned in block 355 is run on the same or a new full slide image. The machine learning model can output a predicted label and a corresponding confidence metric corresponding to each of a plurality of point locations in the image or each of a plurality of portions in the region. The confidence metric can be used to generate a heat map (also in block 365) corresponding to the entire slide image or a portion of the entire slide image, the heat map representing the likelihood (e.g., confidence metric) of the predicted label across different portions of the depicted region. For example, the heat map may be generated by using a smoothing technique or by calculating statistics for each of a plurality of regions.

[0090] At block 370, it is determined whether the machine learning model is fully trained. If so, at block 372, the model is saved along with all the ground truth labels used to train the model.

[0091] Otherwise, the certainty heatmap is displayed (e.g., on the user device), as are the results (predicted labels) at 375 generated by the machine learning model (from block 365). In some cases, the heatmap and the results are displayed simultaneously. For example, a visual indicator of each prediction may be overlaid on the heatmap at a location associated with the corresponding feature of interest. As another example, the heatmap (e.g., of the entire slide image, or for a given field of view) may be presented simultaneously in a non-overlapping manner (e.g., in two frames) with the results. In some cases, the heatmap is presented over a first set of augmented levels, and the results are presented over a second set of augmented levels (which may or may not overlap with the second set of augmented levels).

[0092] At block 380, a new view is selected (e.g., in response to detecting a corresponding input on a user device on which the heatmap was being displayed). At block 385, one or more features of interest are labeled, thereby creating one or more additional ground truth labels (gt). Thus, as shown in block 390, a full set of ground truth labels (G) can be defined to include the ground truth labels from the initialization operation 305 and each additional ground truth label identified via the active learning operation 310.

[0093] At block 395, the machine learning model is retrained using the full set of ground truth labels. At block 397, the full set of ground truth labels is stored (e.g., replacing any previously stored full set of ground truth labels) and / or newly obtained labels (from block 385) are stored to supplement the previously stored ground truth labels. The model can continue to be trained until it is determined (at block 370) that a stopping criterion has been met.

[0094] IV.B. Exemplary Workflows for Using the Digital Pathology Process Flow 4 illustrates an example workflow for using some or all of the digital pathology process flow (e.g., process flow 200 shown in FIG. 2). The illustrated workflow can be executed on one or more computing systems (e.g., a computing system associated with a human annotator, a computing system associated with a user requesting label predictions, and / or a remote computing system (e.g., operating in the cloud)).

[0095] At block 405, some or all of the images (e.g., some or all of the entire slide image) are annotated. Annotation may include (for example) assigning labels (e.g., binary labels, classification labels, numerical labels, etc.) to portions of the images and / or identifying specific portions of the images that are predicted to depict a given feature. Annotation may be performed according to input received at a device operated by the annotator. Annotation may be performed iteratively across different images and / or different portions of interest. Annotation may additionally or alternatively be performed to label different features (e.g., to initially locate individual cells and classify individual cells separately).

[0096] In some cases, annotation is performed at block 215 of process flow 200. In some cases, the annotated images may be stored in image data store 240 such that each image is stored in association with corresponding label data generated through annotation.

[0097] At block 410, data collection is performed. Data collection can include identifying specific annotated images and / or specific portions of annotated images. The specific images and / or specific portions may be annotated to indicate that they have been annotated to identify any specific type of feature of interest, that they have exhibited at least a portion of a specific type of feature of interest, that they have exhibited at least one specific type of feature of interest, and / or that they have exhibited at least a threshold amount of a specific type of feature of interest. For example, data collection can be performed to identify each image patch annotated to indicate that at least 10 cells are exhibited and at least 2 macrophages are exhibited. As another example, data collection can be performed to identify each image patch annotated to identify each macrophage cell (regardless of whether macrophages were detected).

[0098] In some cases, data collection may be performed in the pre-processing stage 210 by sending a query to the image data store 240 and receiving results identifying corresponding images and / or image portions. In some cases, data collection is performed as part of the labeling stage 215.

[0099] At block 415, data preparation is performed. Data preparation can include preprocessing (for example) by one or more preprocessing techniques disclosed herein. Data preparation can include defining one or more tiles or patches (e.g., each corresponding to a portion of the image). Data preparation can include defining a ground truth mask that can identify a portion of the image that depicts at least a portion of the tissue slice. Data preparation can be performed in the preprocessing stage 210.

[0100] In block 420, the prepared data (which may include a plurality of images and at least one label corresponding to each image) may be used to train and / or fine-tune a machine learning model. Training may include using a current version of the machine learning model to generate predictions of the labels (e.g., and corresponding confidence metrics), comparing the predicted labels to the true labels (e.g., using a loss function), and adjusting parameters of the model based on the comparison. Training may further or additionally include fitting the current model using the model and / or one or more static variables (e.g., image features). Training or fine-tuning may be performed in a model training stage 220.

[0101] At block 425, the trained model may be evaluated. The evaluation may characterize (for example) the sensitivity, specificity, and / or accuracy of the predicted labels. The evaluation may be performed in the model training stage 220. The evaluation may include determining whether a given metric exceeds a predefined threshold. If the metric does not exceed a predefined threshold, the training (or fine-tuning) and evaluation may be repeated until the metric exceeds a predefined threshold or until another condition is met (e.g., at least a predefined number of training iterations are completed).

[0102] In block 430, whole slide image inference is performed. Whole slide image inference may be performed in stage 225 and may include generating statistics for each of a plurality of portions of the whole slide image based on a confidence metric associated with the portion of the image within the portion. In some cases, the inference further or alternatively generates statistics for each of a plurality of portions of the whole slide image based on a predicted label associated with the portion of the image within the portion. Thus, the statistics may represent the confidence of the label within the image portion and / or may reflect the overall predicted label of the image portion. Block 430 may be performed by the result generation stage 225 or the (active) model training stage 220.

[0103] In block 435, a confidence heatmap is generated (e.g., in the result generation stage 225 or the (active) model training stage 220) that shows confidence metric statistics for each of a number of parts within the whole slide image. This image can be useful to guide a human annotator as to which parts of the image review or labels are required. For example, red areas can indicate that review of the predicted labels is highly prioritized compared to blue areas.

[0104] At block 440, a new label is identified (e.g., based on input received at the annotator's device). The new label may correspond to the label provided in association with the image portion for which the unreliable prediction was generated.

[0105] The workflow can then return to block 420 and continue training based on the new labels. The iterations can continue until a stopping condition is met (e.g., at least a predetermined number of labels have been provided by annotation, at least a predetermined number of training iterations have been performed, at least a predetermined model accuracy has been achieved, input from a human operator corresponds to acceptance of the model, etc.). V. An Exemplary Framework for Digital Pathology Processing

[0106] As described herein, each of the one or more systems may be involved in performing the various operations and / or processes disclosed for training or using a machine learning model to process images. Figure 5 discloses a network 500 having exemplary systems and applications that may be used in this regard. The network 500 may include components that support some or all of the process flow 200, some or all of the active learning workflow illustrated in Figure 3, and / or some or all of the exemplary workflow illustrated in Figure 4.

[0107] The systems and / or applications may individually or collectively support high-resolution and / or high-speed image storage and retrieval, image visualization, data visualization, interactive data correction, and / or image analysis (e.g., image training and inference).

[0108] Network 500 includes a front-end application sub-network 505 supported by a set of interaction assistance components 510. User devices 512 can communicate (e.g., send communications and receive communications) with a content management system 515, which can manage one or more active learning sessions with user devices 512. Managing the active learning sessions can include coordinating operations performed by one or more other components in network 500.

[0109] The viewer component 520 can define and provide an interface that is presented to the user device 512. The interface can include (for example) an image, one or more input components for labeling the image, one or more predicted labels, or other information as described herein.

[0110] The viewer component 520 can access images by sending a request to the image manager component 530. The request can include (for example) an indication of the type of image being requested (e.g., identifying the organ, stain type, subject's disease type, subject's age, etc.), an identifier for the particular image (e.g., identifying the particular subject), authorization information (e.g., identifying the user associated with the user device 512 to allow the image manager component 530 to evaluate access privileges), identification of a particular clinical trial, and / or an identification of a particular machine learning model being trained or used (which can be associated with a particular image type and / or particular image access restrictions). In some cases, the request is for a single image. In some cases, the request is for multiple images (which the viewer component 520 can present in part or in whole, simultaneously or separately).

[0111] The image management component 530 can evaluate the permissions associated with the request and, if the request is determined to be permitted, can query a remote or local image file system 530 for the image. In some cases, the image management component 530 pre-processes the image before returning it to the viewer component 520. For example, the image management component 530 can crop or scale the image.

[0112] Additionally or alternatively, the viewer component 520 can communicate with the analysis platform component 535 to coordinate image pre-processing. For example, the pre-processing can include identifying a point location of each of one or more cells depicted in the image. The analysis platform component 535 can further facilitate training and / or use of machine learning models. For example, the identified labels based on input from the user device 512 (e.g., predicting whether a cell is a particular cell type, whether the image depicts an artifact, where the tumor border is, etc.) can be used to train a machine learning model. The machine learning model can then be used to generate such predictions and potentially corresponding reliability metrics. The training and use of the machine learning model can be performed using a high-performance component 540 including (for example) a high-speed processor and a large memory. The high-performance component 540 can include multiple nodes 545, each of which can support one or more coding platforms. In some cases, each node 545 operates substantially independently to support parallel processing.

[0113] The analytics platform 535 can detect predicted labels associated with low confidence metrics and can facilitate requests for label review of these predicted labels (e.g., via the content management system 515 and / or the viewer component 520). As described further herein, feedback from the user device 512 indicating confirmation, rejection, or replacement of each label can be used to retrain or fine-tune the model.

[0114] VI. Working Examples A machine learning model was configured to detect the depiction of macrophages in digital pathology images. For each image, a pre-processing step was performed to detect the location of each represented nucleus.

[0115] FIG. 6A depicts an exemplary whole slide image of a stained slice of a sample. Each white box corresponds to a different field of view that the user can enlarge by clicking on the box. Each of the depictions in FIG. 6B-6D corresponds to the same specific field of view. Each of the dots represents the predicted location of a depicted nucleus (identified in preprocessing).

[0116] In Figure 6B, each representation of a nucleus is identical to the others. The interface shown in Figure 6B is presented to the user during the initial training of the machine learning model, allowing the user to select a subset of nuclei to label as macrophages or not. These labels (and others obtained using similar techniques) were used to initially train the model.

[0117] The model then predicted, for each identified nucleus, whether the nucleus corresponds to a macrophage or other cell type. For each nucleus, the model further outputs a confidence metric indicating the reliability of the prediction. Figure 6C shows the same field of view and nuclear identification, but the size and brightness of the dots associated with the nucleus detection are scaled based on the confidence metric. (Specifically, nuclei corresponding to confidence metrics above a threshold are represented by crosses, others by black circles.) These visual characteristics can guide the user to label uncertain cells (e.g., by clicking on the bright dots and providing further input) and / or review the labels of uncertain cells.

[0118] Upon receiving a first round of inputs that verified or corrected the various label predictions, the model was further trained using additional labels. The model then regenerated predictions regarding whether various cells were macrophages and regenerated the prediction and confidence metrics.

[0119] Figure 6D shows the same field of view and nuclear identification, but the size and brightness of the dots associated with nuclear detection are scaled based on a new confidence metric. These updated visual characteristics can guide the user to label or review parts of the image where the updated model is least certain.

[0120] VII. Further Considerations Some embodiments of the present disclosure include a system including one or more data processors. In some embodiments, the system includes a non-transitory computer-readable storage medium including instructions that, when executed on the one or more data processors, cause the one or more data processors to perform some or all of one or more methods and / or some or all of one or more processes disclosed herein. Some embodiments of the present disclosure include a computer program product tangibly embodied in a non-transitory machine-readable storage medium including instructions configured to cause the one or more data processors to perform some or all of one or more methods and / or some or all of one or more processes disclosed herein.

[0121] The terms and expressions used are used as terms of description and not of limitation, and in the use of such terms and expressions there is no intention to exclude equivalents of any features shown and described or portions thereof, but it is recognized that various modifications are possible within the scope of the invention as claimed. Thus, although the invention as claimed has been specifically disclosed by embodiments and optional features, it is to be understood that modifications and variations of the concepts disclosed herein may be resorted to by those skilled in the art, and such modifications and variations are considered to be within the scope of the invention as defined by the appended claims.

[0122] This description provides only preferred exemplary embodiments and is not intended to limit the scope, applicability, or configuration of the present disclosure. Rather, this description of exemplary embodiments provides those skilled in the art with an enabling description for implementing various embodiments. It will be understood that various changes can be made in the function and arrangement of elements without departing from the spirit and scope of the appended claims.

[0123] Specific details are given herein to provide a thorough understanding of the embodiments. However, it will be understood that the embodiments can be practiced without these specific details. For example, circuits, systems, networks, processes, and other components may be shown as components in block diagram form so as not to obscure the embodiments in unnecessary detail. In other examples, well-known circuits, processes, algorithms, structures, and techniques may be shown without unnecessary detail so as to avoid obscuring the embodiments.

Claims

1. 1. A method for active learning in a digital pathology system, comprising: accessing a machine learning model configured to process the image using the one or more parameters to generate a label; executing the machine learning model, converting at least a portion of each of the at least one digital pathology image into a plurality of predictive labels; generating a confidence metric for each of the plurality of predicted labels; and executing the machine learning model such that generating, for each of the at least one digital pathology image, a confidence map for the digital pathology image; the certainty map comprises a set of values, each value of the set of values ​​(i) corresponding to a portion of the digital pathology image and (ii) identifying a statistical value of one or more confidence metrics corresponding to a point location or region within the portion; Generating, utilizing an interface to render the at least a portion of each of the at least one digital pathology image to differentially represent predicted labels based on a corresponding confidence metric; utilizing a representation of the certainty map based on the set of values. Utilizing said interface; receiving a label input that confirms, rejects, or replaces at least one of the plurality of predicted labels in response to utilization of the interface; updating the one or more parameters of the machine learning model based on the label input; A method comprising:

2. The method of claim 1 , wherein a predictive label from the plurality of predictive labels identifies a classification of cells and / or an estimated amount of a particular type of cell within the at least a portion of the at least one digital pathology image.

3. The interface, receiving user input representing a statistic relating to a reliability metric at one zoom level; represents the confidence metric at another level of augmentation, receiving a user input to change the magnification level; The method according to claim 1 or 2, wherein the method is configured to:

4. receiving a new digital pathology image; converting the new digital pathology image into one or more new labels that characterize the new digital pathology image or a portion of the new digital pathology image; outputting said one or more new labels; The method of any one of claims 1 to 3, further comprising:

5. pre-processing the initial version of the digital pathology image to detect a location of each feature of interest; 5. The method of claim 1, wherein transforming the at least a portion of the digital pathology image comprises generating a label corresponding to each detected location of interest within the at least a portion of the digital pathology image.

6. for each of the plurality of predicted labels, selecting a size, color, or shape of a marker based on the confidence metric of the predicted label; 6. The method of claim 1, wherein the interface represents each predicted label of the plurality of predicted labels using a marker having the selected size, color, or shape, thereby differentially representing predicted labels based on a corresponding confidence metric.

7. The method of any one of claims 1 to 6, wherein the machine learning model comprises a deep neural network.

8. 1. A system comprising: one or more data processors; a non-transitory computer-readable storage medium containing instructions that, when executed on the one or more data processors, cause the one or more data processors to perform a set of operations; The set of operations is accessing a machine learning model configured to process the image using the one or more parameters to generate a label; executing the machine learning model, converting at least a portion of each of the at least one digital pathology image into a plurality of predictive labels; generating a confidence metric for each of the plurality of predicted labels; and executing the machine learning model such that generating, for each of the at least one digital pathology image, a confidence map for the digital pathology image; the certainty map comprises a set of values, each value of the set of values ​​(i) corresponding to a portion of the digital pathology image and (ii) identifying a statistical value of one or more confidence metrics corresponding to a point location or region within the portion; Generating, utilizing an interface to render the at least a portion of each of the at least one digital pathology image to differentially represent predicted labels based on a corresponding confidence metric; utilizing a representation of the certainty map based on the set of values. Utilizing said interface; receiving a label input that confirms, rejects, or replaces at least one of the plurality of predicted labels in response to utilization of the interface; updating the one or more parameters of the machine learning model based on the label input; Including, the system.

9. The system of claim 8 , wherein a predictive label of the plurality of predictive labels identifies a classification of cells and / or an estimated amount of a particular type of cell within the at least a portion of the at least one digital pathology image.

10. The interface, receiving user input representing a statistic relating to a reliability metric at one zoom level; represents the confidence metric at another level of augmentation, receiving a user input to change the magnification level; The system according to claim 8 or claim 9, configured to:

11. The set of operations is receiving new digital pathology images; and converting the new digital pathology image into one or more new labels that characterize the new digital pathology image or a portion of the new digital pathology image; outputting said one or more new labels; The system according to any one of claims 8 to 10, further comprising:

12. The set of operations is Preprocessing an initial version of the digital pathology image to detect a location of each feature of interest. and wherein transforming the at least a portion of the digital pathology image comprises generating a label corresponding to each detected location of interest within the at least a portion of the digital pathology image.

13. The set of operations is For each of the plurality of predicted labels, selecting a size, color, or shape of a marker based on the confidence metric of the predicted label.

12. The system of claim 8, further comprising: wherein the interface uses a marker having the selected size, color, or shape to represent each predicted label of the plurality of predicted labels using a marker, thereby differentially representing predicted labels based on a corresponding confidence metric.

14. The system of any one of claims 8 to 13, wherein the machine learning model comprises a deep neural network.

15. A non-transitory machine-readable storage medium having instructions configured to cause one or more data processors to perform a set of operations, comprising: The set of operations is accessing a machine learning model configured to process the image using the one or more parameters to generate a label; executing the machine learning model, converting at least a portion of each of the at least one digital pathology image into a plurality of predictive labels; generating a confidence metric for each of the plurality of predicted labels; and executing the machine learning model such that generating, for each of the at least one digital pathology image, a confidence map for the digital pathology image; the certainty map comprises a set of values; each value in the set of values ​​(i) corresponds to a portion of the digital pathology image; and (ii) identifies a statistical value of one or more confidence metrics corresponding to a point location or region within the portion. Generating, utilizing an interface to render the at least a portion of each of the at least one digital pathology image to differentially represent predicted labels based on a corresponding confidence metric; utilizing a representation of the certainty map based on the set of values. Utilizing said interface; receiving a label input that confirms, rejects, or replaces at least one of the plurality of predicted labels in response to utilization of the interface; updating the one or more parameters of the machine learning model based on the label input; A non-transitory machine-readable storage medium comprising:

16. The non-transitory machine-readable storage medium of claim 15, wherein a predictive label among the plurality of predictive labels identifies a classification of cells and / or an estimated amount of a particular type of cell within the at least a portion of the at least one digital pathology image.

17. The interface, receiving user input representing a statistic relating to a reliability metric at one zoom level; represents the confidence metric at another level of augmentation, receiving a user input to change the magnification level; 17. The non-transitory machine-readable storage medium of claim 15 or 16, configured to:

18. The set of operations is receiving new digital pathology images; and converting the new digital pathology image into one or more new labels that characterize the new digital pathology image or a portion of the new digital pathology image; outputting said one or more new labels; The non-transitory machine-readable storage medium of any one of claims 15 to 17, further comprising:

19. The set of operations is pre-processing the initial version of the digital pathology image to detect a location of each feature of interest; 19. The non-transitory machine-readable storage medium of claim 15, wherein transforming the at least a portion of the digital pathology image includes generating a label corresponding to each detected location of interest within the at least a portion of the digital pathology image.

20. The set of operations is for each of the plurality of predicted labels, selecting a size, color, or shape of a marker based on the confidence metric of the predicted label; 20. The non-transitory machine-readable storage medium of claim 15, wherein the interface differentially represents predicted labels based on a corresponding confidence metric by representing each predicted label of the plurality of predicted labels using a marker with the marker having the selected size, color, or shape.

Images