Blast cell classification

Abstract

A computer-implemented method for distinguishing between lymphoid blast cells and myeloblast cells includes receiving a digital image including one or more blast cells (S30) and applying a parametric model classifier (412) to one or more portions of the digital image, each including a respective blast cell (S34), where the parametric model (416) is configured to generate an output indicating whether each blast cell is a lymphoblast cell or a myeloblast cell (S38). A computer-implemented method for training the parametric model (416) is also provided, as well as a clinical decision support system (400) that relies on the computer-implemented method for classifying blast cells.

Description

[Technical field]

[0001] The present invention relates to a computer-implemented method for distinguishing between myeloid and lymphoblast cells and an associated method for diagnosing acute myeloid or acute lymphoblastic leukemia based on the output of the discrimination. A computer-implemented method for training a deep neural network, as well as a clinical decision support system, are also provided. [Background technology]

[0002] Blast cells are precursors of mature blood cells that are found circulating in the human bloodstream. Normally, blast cells are confined to a person's bone marrow. However, when a patient suffers from leukemia, abnormal blast cells grow uncontrollably in the bone marrow to such an extent that the production of other cells important for survival is prevented. Furthermore, the uncontrolled growth also causes the abnormal blast cells to leak into the person's bloodstream. Thus, leukemia can be diagnosed by the detection of these abnormal blast cells in the patient's bloodstream.

[0003] Acute leukemia manifests itself in forms including acute myeloid leukemia (AML) and acute lymphoblastic leukemia (ALL), each of which has several subtypes. To determine what type of leukemia is present, it is necessary to classify the abnormal blast cells as either myeloblasts or lymphoblasts. Up until this point, this classification has been very difficult because the cells are immature and lack lineage discrimination.

[0004] Current methods for classifying abnormal blast cells as myeloblasts or lymphoblasts rely on several stages of discrimination cluster (CD) marker evaluation. An example of a workflow that may be used is as follows: In the first step, a complete blood count (CBC) is obtained as part of routine screening or due to the manifestation of possible symptoms of leukemia. If the CBC results show an abnormal result (i.e., an abnormal number of blast cells in the blood), a blood smear may be taken and examined by a hematologist. If, on analysis of the blood smear, it is confirmed that abnormal blast cells are present in the blood, a repeat sample may be taken for confirmation. Further analysis may then be performed, including CD marker evaluation. CD marker evaluation is used to determine the cell lineage (i.e., whether the blast cells are myeloblasts or lymphoblasts). After that determination is made, a CD marker panel of specific myeloid or lymphoid cell lines may be performed, ultimately leading to a diagnosis. It may be appreciated that this is a lengthy process requiring several stages of CD marker evaluation to arrive at a diagnosis. In combination with CD marker assessment, other techniques may be used, including analysis of cerebrospinal fluid (CSF) or bone marrow samples using cytogenetics, fluorescent in situ hybridization (FISH) or polymerase chain reaction (PCR) techniques. These processes are similarly time-consuming and often expensive. Summary of the Invention

[0005] The present invention aims to address this by providing a computer-implemented method for distinguishing between myeloblast and lymphoblast cells.

[0006] At a high level, the invention provides a computer-implemented method that uses a parametric model to distinguish between lymphoblast cells and myeloblast cells, or to determine whether blast cells in a digital image are of myeloid or lymphoid lineage origin.

[0007] More specifically, a first aspect of the present invention provides a computer-implemented method for distinguishing between lymphoblast cells and myeloblast cells, the computer-implemented method comprising receiving a digital image comprising one or more blast cells, and applying a parametric model classifier to one or more portions of the digital image, each portion comprising a respective blast cell, the parametric model classifier being configured to generate an output indicative of whether each blast cell is a lymphoblast cell or a myeloblast cell.

[0008] By using a classifier at an early stage, as required by the computer-implemented method according to the first aspect of the invention, it is possible to reduce or avoid the need for various stages of CD marker analysis (or other lengthy techniques) to arrive at a definitive diagnosis of AML or ALL.

[0009] Optional features of the first aspect of the invention are described herein. It should be noted that any or all of these features may be combined with each other unless the context indicates otherwise or it is expressly stated that a particular feature is incompatible or cannot be combined with another feature.

[0010] The digital image may be received from an imaging device. In addition to the computer-implemented steps, the method according to the first aspect of the present invention may also include capturing a digital image of one or more blast cells using an imaging device. The imaging device may comprise a camera. The imaging device may further comprise a microscope. Thus, the digital image may be a microscopic image, such as a bright-field microscopic image. However, alternatively, the digital image may be based on phase contrast imaging, differential interference contrast microscopy, or dark-field microscopy.

[0011] The digital image of the one or more blast cells may be a digital image of a sample on a slide.

[0012] In addition to the computer-implemented steps, the method according to the first aspect of the invention may further comprise preparing a slide containing one or more blast cells. Preferably, the slide is prepared using a method in which a drop of blood dries in the presence of air or with a moderate air flow, and a technique that creates a monolayer of all cells in the volume transferred to the slide. Again, preferably, the slide is prepared in such a way that all single cells can be counted and identified by type. To ensure this, a suitable diluent, dilution factor, means must be selected that allow the droplet to be spread over a relatively large area by improving the hydrophilicity of the glass slide and / or by mechanically spreading the droplet. The appropriate choice of parameters and methods results in a very thin liquid layer that dries quickly so that the presentation and preservation of the cells mimics that found in a conventional blood "smear". Details of how advantageous slides can be prepared can be found in WO 2012 / 030313, which is incorporated by reference in its entirety. Further disclosure regarding slide preparation can be found in U.S. Patent Application Publication No. 2009 / 0269799, U.S. Patent Application Publication No. 2010 / 284602, U.S. Patent Application Publication No. 2011 / 014645, and U.S. Patent Application Publication No. 2016 / 209320, each of which is incorporated by reference in its entirety.

[0013] As can be seen from the above references, the sample is preferably stained. Thus, the digital image can be an image of a stained sample on a slide. Stains that can be used include the following: Romanowsky stain, Giemsa stain, Jenner stain, Wright stain, Field stain, May-Grunwald stain, Leishman stain. It can be seen that other types of stains can also be used.

[0014] The computer-implemented method may further include, after receiving a digital image including one or more blast cells, identifying one or more blast cells in the digital image. In other words, before applying the classifier to each blast cell, the computer-implemented method may further include identifying where the blast cells are located in the image. Even in patients suffering from AML or ALL, the proportion of blast cells in blood is still very low, for example compared to red blood cells. Therefore, it is beneficial to identify blast cells in the digital image before applying the classifier, so that the classifier acts only on the identified blast cells. Identifying blast cells may include applying an image analysis algorithm to the digital image. The image analysis algorithm is preferably configured to identify a bounding box around one or more blast cells in the digital image. In particular, the image analysis algorithm is preferably configured to identify a respective bounding box around each of the one or more blast cells in the digital image. In this specification, identifying a bounding box should be understood to mean identifying an area, preferably a square or rectangular area, that preferably includes a single blast cell. In some cases, the image analysis algorithm may be configured to generate a plurality of image files, each image file including a digital image of a blast cell of one or more blast cells. The boundary of the image in each image file of the plurality of image files may be a bounding box as described above. In some cases, the plurality of files may be generated after the image analysis algorithm is applied to the digital image. If a plurality of image files are generated, the computer-implemented method may include applying a parametric model classifier to each of the generated image files, the parametric model classifier configured to generate a respective output indicative of whether each blast cell in each image is a lymphoblast cell or a myeloblast cell. Thus, the output of the computer-implemented method may include a plurality of outputs, each indicative of whether each blast cell of the one or more blast cells is a lymphoblast cell or a myeloblast cell.

[0015] We now consider in more detail the form of the output of the parametric model classifier. For each blast cell in the digital image or for each image file, the output of the parametric model classifier may include a numerical value x indicating whether the blast cell is a lymphoblast cell or a myeloblast cell. In some cases, the output (again for each blast cell in the digital image or for each image file) may include a first value and a second value, the first value indicating the likelihood that the blast cell is a lymphoblast cell (or that the image file contains an image of a lymphoblast cell) and the second value indicating the likelihood that the blast cell is a myeloblast cell (or that the image file contains an image of a myeloblast cell). Preferably, the first and second values ​​are probabilities, preferably summing to 1. The more extreme the values, the more confident the result. In other words, the numerical value x may be in the range of [0,1], and if the value x is equal to 1, the blast cells may be identified as lymphoblast cells with 100% confidence, and if the value x is equal to 0, the blast cells may be identified as a type of cell other than lymphoblast cells with 100% confidence. Alternatively, the numerical value x may be in the range of [0,1], and if the value x is equal to 1, the blast cells may be identified as myeloblast cells with 100% confidence, and if the value x is equal to 0, the blast cells may be identified as a type of cell other than myeloblast cells with 100% confidence. It is known that distinguishing between myeloblast cells and lymphoblast cells may be difficult, and therefore, it is conceivable that in some cases (either due to the nature of the blast cells themselves or due to, for example, the angle at which it is shown in the digital image) it is not possible to accurately classify the cells. Thus, in some cases, if the numerical value is within a predetermined range, the output of the parametric classifier is configured to be indeterminate or indeterminate. Such results may be discarded. The predetermined value may be between 0.1 and 0.9, between 0.2 and 0.8, between 0.3 and 0.7, between 0.4 and 0.6, or between 0.45 and 0.55.

[0016] Each output may be of the form: [first value, second value]. Consider the case where the predefined threshold range is from 0.4 to 0.6. - If the result is [0.98,0.02], it can be concluded with 98% certainty that the blast cells in question are lymphoblast cells. -If the result is [0.43,0.57], the result is considered indeterminate since the value falls within the range of 0.4 to 0.6. - If the result is [0.22,0.78], it can be concluded with 78% certainty that the blast cells in question are myeloblast cells.

[0017] In these cases, the blast cells are assumed to be either myeloblasts or lymphoblasts.

[0018] Throughout the present application, reference is made to "parametric model classifiers". In the context of this application, a parametric model is one that assumes a parametric form of a function for generating an output from input data (i.e. data representing a digital image), the function comprising a fixed number of parameters. The parametric form, as the name suggests, depends on multiple parameters, the goal of which, for example, in a training process, is to identify those parameters. In contrast, non-parametric models are unbounded and have no limit to their complexity. Advantages of parametric model classifiers (compared to non-parametric classifiers) include simplicity, speed, and the fact that they can produce reliable results with smaller amounts of data. The parametric model classifier is preferably a machine learning based classifier. In a preferred case, the classifier is based on a convolutional neural network, such as a deep neural network. One example of a suitable neural network classifier is the residual neural network classifier. 123(herein "ResNet"). ResNet is an artificial neural network based on constructs known from pyramidal cells of the cerebral cortex. ResNet works on the principle of skip connections or shortcuts to jump layers in a neural network. Typical ResNet models are implemented with two or three layers skipping, including nonlinearities, or batch normalization in between. There are two reasons why connections can be skipped: to avoid the problem of vanishing gradients, or to mitigate the problem of degradation (precision saturation), where adding more layers to a suitably deep model results in higher training error. During training, weights are adapted to mute upstream layers and amplify previously skipped layers. In the simplest case, only the weights of the connections of adjacent layers are adapted, and there are no explicit weights of the upstream layers. This works best if a single nonlinear layer is stepped over, or if the intermediate layers are all linear. If not, an explicit weight matrix can be learned for the skipped connections. ResNet is advantageous because by skipping, it effectively simplifies the network by using fewer layers in the initial training phase, thereby speeding up learning by reducing the effect of vanishing gradients since there are fewer layers to propagate. The network then gradually restores the skipped layers as it learns the feature space. Towards the end of training, when all layers are expanded, it stays close to the manifold and therefore learns faster. Examples of suitable ResNets include ResNet34, ResNet50, and ResNet100, where the numbers represent the number of layers present in the residual neural network. In our implementation, we use Inception 4 , VGG 5 , and EfficientNet 6 Other types of convolutional neural networks such as may also be used.

[0019] Alternatively or additionally, the model may be based on Ilse et al. (2018) 7 For example, deep attention MIL can be created using multiple instance learning, as described in.

[0020] Rather than neural network methods that involve training a model using per-image labels, a sham model can be used. A sham model takes two images and must determine if they are of the same class. Using labels applied to the full slide or to each cell individually, each pair can be identified as true if it consists of the same cell type (e.g., two lymphoblast cells or two myeloblast cells), or as false if the images are from cells of different types.

[0021] The objective of the present invention is to improve the efficiency with which a patient can be diagnosed as either AML or ALL (or neither). Thus, we now consider how the output of the parametric model classifier is used. Thus, the computer-implemented method may further include calculating a patient-level score based on the output value x of each of all blast cells in the digital image. In this way, a diagnosis or clinical decision may be made based on multiple blast cells rather than a single blast cell. The patient-level score may include one or more of the mean, median, maximum, or minimum value. Specifically, a patient-level score may be calculated for all of the x values ​​representing the likelihood that a blast cell is a lymphoblast cell and / or the likelihood that a blast cell is a myeloblast cell. If the numerical value of the patient-level score is within a predetermined range, the output of the parametric classifier is configured to be indeterminate or indeterminate. Such results may be discarded.

[0022] In some cases, the computer-implemented method of the first aspect of the present invention may be used as part of a decision support system that allows a clinician to select an appropriate course of action. In such cases, it may be desirable for the clinician to review the results generated by the parametric model classifier. Specifically, the computer implementation may further include generating instructions based on the output of the parametric model classifier, configured to cause a display device of the computing system to display a gallery, the gallery including a first plurality of images showing blast cells identified with the highest confidence as lymphoblast cells and a second plurality of images showing blast cells identified with the highest confidence as myeloblast cells. In this context, "with the highest confidence" may be understood to mean blast cells that have the highest probability of being a particular type of blast cell (i.e., myeloblast cells or lymphoblast cells). This allows the clinician to review the images and determine an appropriate course of action to determine the patient's prognosis. In some cases, the first plurality of images includes the same number of images as the second plurality of images. This number may be between 1 and 100, more preferably between 5 and 50, more preferably between 10 and 25, and most preferably about 20.

[0023] As discussed above in this application, it is often necessary to perform CD marker screening to arrive at a more definitive diagnosis. Various types of CD marker screening are available, each targeting a specific biomarker. In some cases, the clinician may identify the CD marker to be used for screening based on the output of the parametric model classifier. However, in other cases, the computer-implemented method may further include selecting one or more CD markers from a plurality of available CD markers based on the output of the parametric model classifier. In this way, the CD marker to be used for subsequent testing can be automatically identified based on the output of the parametric model.

[0024] For a parametric model to provide accurate results, it is often necessary to train it using known data. Thus, a second aspect of the present invention provides a computer-implemented method for generating a parametric model classifier configured to distinguish lymphoblast cells and myeloblast cells in a digital image, the computer-implemented method comprising: receiving a plurality of pairs of labeled training data, each pair of labeled training data comprising input data comprising a digital image of a blast cell from a patient diagnosed with either acute myeloid leukemia or acute lymphoblastic leukemia, and output data comprising an indication of whether the patient has acute myeloid leukemia or acute lymphoblastic leukemia; and training the parametric model classifier using the training data. Any features described above with respect to the first aspect of the present invention apply equally well to the second aspect of the present invention, where applicable. For the avoidance of doubt, it is emphasized that in the second aspect of the present invention, the parametric model classifier may be a machine learning based classifier. In a preferred case, the classifier is based on a convolutional neural network, such as a deep neural network. One example of a suitable neural network classifier is a residual neural network classifier (herein, "ResNet"). Examples of suitable ResNets include ResNet34, ResNet50, and ResNet100, where the numbers represent the number of layers present in the residual neural network. Convolutional neural networks are classified using the Adam optimizer 8 Convolutional neural networks can be trained using the Smith 9 A framework that can be used to train convolutional neural networks is the fastai 10 and Paszke et al. 11 A library of 3D neural networks can be used. A different learning strategy that relies instead on a flat learning rate and cosine annealing can also be used. Empirical models such as visual transformers or capsules can also be used rather than convolutional neural networks.

[0025] The indication of whether the patient has acute myeloid leukemia or acute lymphocytic leukemia may include a numerical value. Specifically, if the patient has acute myeloid leukemia, the numerical value may be 1, and if not, the numerical value may be 0. Conversely, if the patient has acute lymphocytic leukemia, the numerical value may be 1, and if not, the numerical value may be 0. More detailed information regarding the training process is provided in the "Experimental Results" section later in this application.

[0026] In the computer implemented method of the first aspect of the invention, the parametric model classifier may be trained using the computer implemented method of the second aspect of the invention.

[0027] A third aspect of the present invention provides a computer-implemented method for providing a provisional diagnosis of acute myeloid leukemia or acute lymphocytic leukemia, the computer-implemented method comprising: executing the computer-implemented method of the first aspect of the present invention; and determining whether a patient whose blast cells are shown in the digital image suffers from acute myeloid leukemia, acute myeloid leukemia, or neither based on the output of the parametric model classifier or a patient-level score calculated based on the output of the parametric model classifier. As mentioned above, all optional features described with respect to the first aspect of the present invention or the second aspect of the present invention apply equally well to the computer-implemented invention of the third aspect of the present invention. Furthermore, based on the result of the determination, the computer-implemented method of the third aspect of the present invention may further comprise generating instructions configured to cause a display device of the computing system to display the result of the determination.

[0028] The first to third aspects of the invention relate to computer implemented methods. Corresponding fourth to sixth aspects of the invention provide a computer program product which causes a computer to carry out the computer implemented method of each of the first to third aspects of the invention when the program is executed by a computer or other computing device. The seventh to ninth aspects of the invention provide a computer readable data carrier storing the computer program product of the fourth to sixth aspects of the invention, respectively.

[0029] As explained above, it is an object of the present invention to provide a clinical decision support system for assisting clinicians in diagnosing acute myeloid leukemia and / or acute lymphoblastic leukemia. Accordingly, a tenth aspect of the present invention provides a clinical decision support system comprising a computing device having a processor, the processor configured to perform the computer-implemented method of any one of the first to third aspects of the present invention.

[0030] An eleventh particularly preferred aspect of the invention provides a computer implemented method for distinguishing between lymphoblast cells and myeloblast cells comprising the steps of: receiving a digital image comprising one or more blast cells; applying image analysis to the digital image, the image analysis algorithm being configured to detect one or more blast cells in the digital image and generate a bounding box around each of the one or more blast cells; generating a plurality of image files, each image file comprising a digital image of a blast cell of the one or more blast cells, the boundary of the image in each file corresponding to the respective bounding box; and applying a deep neural network classifier to each of the generated image files, wherein each blast cell is a lymphoblast cell or applying the deep neural network classifier to generate an output indicative of whether the cell is a myeloblast cell, the output comprising a numerical value x in a range [0,1], where if the value x is equal to 1, the blast cell is identified as a lymphoblast cell with 100% confidence, and if the value x is equal to 0, the blast cell is identified as a myeloblast cell with 100% confidence, or if the value x is equal to 1, the blast cell is identified as a myeloblast cell with 100% confidence, and if the value x is equal to 0, the blast cell is identified as a lymphoblast cell with 100% confidence; and calculating a patient-level score for all blast cells in the digital image based on the respective output values ​​x, where the patient-level score comprises a mean, a median, a maximum value of x, and / or a minimum value of x. Any feature described above with respect to the first to tenth aspects of the present invention applies equally well to the eleventh aspect of the present invention, unless technically clearly incompatible or the context clearly indicates otherwise.

[0031] Optionally, the deep neural network classifier of the eleventh aspect of the present invention comprises receiving a plurality of pairs of labeled training data, each pair of labeled training data comprising input data comprising a digital image of a blast cell from a patient diagnosed with either acute myeloid leukemia or acute lymphoblastic leukemia, and output data comprising an indication of whether the patient has acute myeloid leukemia or acute lymphoblastic leukemia, the output comprising a numerical value x, where the output takes on a value of 0 or 1, and where if the value x is equal to 1, then the blast cell is identified as having a blast cell. and training the deep neural network classifier using the training data; and receiving a plurality of pairs of labeled training data, whereby if the value x is equal to 0, a blast cell is identified with 100% confidence as a lymphoblast cell and if the value x is equal to 0, a blast cell is identified with 100% confidence as a myeloblast cell, or if the value x is equal to 0, a blast cell is identified with 100% confidence as a lymphoblast cell and if the value x is equal to 1, a blast cell is identified with 100% confidence as a myeloblast cell.

[0032] A twelfth aspect of the present invention provides a clinical decision support system comprising a computing device having a processor configured to provide a provisional diagnosis of acute myeloid leukemia or acute lymphocytic leukemia by executing the computer-implemented method of the eleventh aspect of the present invention, the processor being further configured to generate instructions configured to determine whether a patient whose blast cells are shown in the digital image is affected by acute myeloid leukemia, acute lymphocytic leukemia, or neither leukemia based on a patient-level score calculated based on an output of the deep neural network classifier, and based on the result of the determination, to cause a display of the computing system to display the result of the determination. [Brief description of the drawings]

[0033] Embodiments of the present invention will now be described with reference to the accompanying drawings.

[0034] [Figure 1]1 illustrates a system that may be used to perform computer-implemented methods according to some aspects of the present invention. [Figure 2A] 2 shows a raw image including blast cells that can be obtained by the imaging device shown in FIG. 1. [Figure 2B] 2 shows a raw image including blast cells that can be obtained by the imaging device shown in FIG. 1. [Diagram 3] 1 is a flow chart showing a method for high level blast cell classification. [Figure 4] FIG. 13 is an example of a diagram that may be generated on a display based on the output of the blast cell classification. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0035] Aspects and embodiments of the present invention will now be described with reference to the accompanying figures. Further aspects and embodiments will be apparent to those skilled in the art. All documents mentioned in this specification are incorporated herein by reference.

[0036] FIG. 1 illustrates a system 100 including various components for performing or enabling the performance of computer-implemented methods according to various aspects of the present invention. The system 100 includes a slide preparation device 200, an imaging device 300, and a clinical decision support system 400. The slide preparation device 200 is preferably configured to prepare a slide including one or more blast cells. As previously discussed in this application, slide preparation may use methods described in WO 2012 / 030313, US 2009 / 0269799, US 2010 / 284602, US 2011 / 014645, and US 2016 / 209320, each of which is incorporated herein by reference in its entirety. Slide preparation is outside the scope of this application and will not be discussed in further detail. Once the slide is prepared, the slide is imaged using an imaging device 300, which may include, for example, a microscope and a camera (not shown). When the slide preparation device 200 prepares the slide as in the above references, a monolayer is formed, which allows individual cells (including red blood cells, white blood cells, platelets, and crucially blast cells) to be visualized and counted. The clinical decision support system 400 then receives an image including blast cells from the imaging device 300. Examples of the types of images that may be received are shown in FIG. 2A and FIG. 2B. The clinical decision support system 400 is then used to identify and classify the blast cells in the image. The example clinical decision support system shown in FIG. 1 includes an imaging device interface module 402, a processor 404, a memory 406, and a display 408. It should be noted that in some alternative configurations, the display 408 may be an external component. In those cases, the processor 404 of the clinical decision support device 400 may further include a display module interface module (not shown). The processor 404 of the clinical decision support system 400 may include a blast cell identification module 410, a blast cell classification module 412, and a graphical user interface (GUI) generation module 414.As used herein, a "module" may be in the form of a physical or functional module implemented, for example, in the form of a software module (i.e., computer readable code). The memory 406 may include a parametric model 416 that may be applied by the blast cell classification module 412 to an image including blast cells. The memory may further include a buffer 418.

[0037] We will now describe the operation of the clinical decision support device 400 for identifying and classifying blast cells and generating instructions for displaying various images on the display 408. Figure 3 is a flow chart illustrating the high level steps of the computer-implemented method.

[0038] In a first step S30, an image containing blast cells is received from the imaging device 300 at the clinical decision support system 400 via its imaging device interface module 402. Then, in step S32, blast cells are identified in the image by the blast cell identification module 410. The output of this process can be, for example, several image files (which may be stored temporarily in a buffer 418 or more permanently in the memory 406), each containing an image of a single blast cell from the image. Alternatively, rather than generating multiple image files, the blast cell identification module 410 may identify blast cells in the image and define the boundaries of regions each containing a single blast cell. The output of the blast cell identification module 410 may be in the form of a list of pixel arrays, each pixel array corresponding to a region of the image containing a single blast cell. This list may also be stored in the buffer 418 or more permanently in the memory 406. In step S34, the parametric model 416 is applied by the blast cell classification module 412 to each image file (or a region of the image containing blast cells) to determine whether the blast cells are lymphoblast cells or myeloblast cells. As mentioned above, the output of the blast cell classification module 412 is a number, preferably between 0 and 1, for each blast cell, which number represents the probability or confidence level that the blast cell in question is either a myeloblast cell or a lymphoblast cell. Alternatively, the model may return two probabilities, namely, the probability that the blast cell is a lymphoblast cell and the probability that the blast cell is a myeloblast cell. These probabilities should be added to 1 (or 100%, or equivalent). Various different steps may then be performed by the blast cell classification module 412 once the multiple probabilities have been calculated using the parametric model 416. In FIG. 3, these are represented as different branches of a flowchart, but it should be emphasized that this should not be understood to mean that the computer-implemented methods according to various aspects of the present invention cannot include two or more of the steps on the different branches. In step S36, the patient score is calculated and stored in memory 406 (eg, in buffer 418).As used herein, the patient-level score is a statistical parameter that represents a probability value calculated for each of the images of (single) blast cells. As mentioned above, the patient-level score may take various forms. In step S38, the GUI generation module 414 may be configured to generate instructions based on the output of the blast cell classification module 412, which, when received by the display 408, causes the display to present the results to a user of the clinical decision support system 400. For example, the display 408 may display a gallery containing the 18 cells most likely to be lymphoblast cells and the 20 cells most likely to be myeloblast cells, as shown in FIG. 4. Of course, the display does not have to be in the form of such a gallery, which is just one option. In some cases, the patient-level score and the individual probability of each blast cell may also be displayed.

[0039] In some examples not shown, the method may further include determining appropriate CD markers for a subsequent screening step, for example based on the patient-level score.

[0040] Experimental Results Having described the computer-implemented method of the present invention, we now present some experimental data that demonstrates the effectiveness of the present invention. Two experiments were performed on different datasets, referred to herein as the Toulouse dataset and the Boston dataset.

[0041] From the following it can be seen that computer implemented methods according to various aspects of the present invention are highly reliable in distinguishing between myeloblast and lymphoblast cells.

[0042] A. Toulouse Dataset The study used images from blood smears of 119 patients, of which 52 had acute myeloid leukemia and 67 had acute lymphoblastic leukemia (51 ALL-B, 16 ALL-T). Images of the blood smears were acquired at high resolution with the CellaVision device.

[0043] We randomly sampled 16 slides as a validation set and used the remainder to train a neural network, in this case a ResNet50 network.

[0044] In the validation set, an AUC (area under the curve in a receiver operated plot) of 83.91% was achieved per cell. After aggregating the results, i.e., by generating patient-level scores, the AUC increased to 95.31%.

[0045] The same model was tested on an additional dataset from the same hospital containing 125 slides from AML / ALL cases with lower blast counts, on which the same model achieved an AUC of 0.89963.

[0046] B. Boston Dataset In this experiment, 39 slides were obtained from AML or ALL patients. Slides were printed and imaged using the methods described earlier in this application. Images were obtained at 20x magnification using a high-resolution camera. Of the 39 slides, 21 were from AML patients and 18 were from ALL patients. A neural network (ResNet50 network) was trained on 29 slides and 10 slides (5 AML and 5 ALL) were kept as a validation set. Depending on which slides were selected for the validation set, the AUC was between 79.51% and 91.55%.

[0047] C. Alternative Training and Evaluation Procedures The dataset is split into training and validation sets by excluding complete slides rather than randomly selecting cells from the complete dataset.

[0048] To thoroughly test per-slide performance for the entire dataset, we performed slide-wise cross-validation where we left out one slide and trained the model on all others. We evaluated performance on the left-out slide. This was repeated for all slides.

[0049] We used an ensemble of five models (also using ResNet34). The ensemble yields better performance as random errors from the individual models are corrected. An overall AUC of 86.09% was achieved over the entire dataset.

[0050] General description of the application The features disclosed in the foregoing description, or in the following claims, or in the accompanying drawings, and expressed in a specific form or in terms of means for performing a disclosed function or a method or process for obtaining a disclosed result, may be utilized, individually or in any combination of such features, as appropriate, to realize the invention in its diverse forms.

[0051] While the present invention has been described in conjunction with the exemplary embodiments set forth above, many equivalent modifications and variations will be apparent to those skilled in the art given this disclosure. Accordingly, the exemplary embodiments of the present invention set forth above are considered to be illustrative and not limiting. Various changes may be made to the described embodiments without departing from the spirit and scope of the present invention.

[0052] For the avoidance of doubt, any theoretical explanations provided herein are provided for the purpose of enhancing the understanding of the reader, and the inventors do not wish to be bound by any of these theoretical explanations.

[0053] Any section headings used herein are for organizational purposes only and are not to be construed as limiting the subject matter described.

[0054] Throughout this specification, including the claims which follow, unless the context requires otherwise, the words "comprise" and "include", as well as variations such as "comprises", "comprising" and "including", are understood to mean the inclusion of a stated integer or step or group of integers or steps but not the exclusion of any other integer or step or group of integers or steps.

[0055] It should be noted that, as used in this specification and the appended claims, the singular forms "a," "an," and "the" include plural referents unless the context clearly dictates otherwise. Ranges may be expressed herein as from "about" one particular value and / or to "about" another particular value. When such a range is expressed, another embodiment includes from the one particular value and / or to the other particular value. Similarly, when values ​​are expressed as approximations, by use of the antecedent "about," it will be understood that the particular value forms another embodiment. The term "about" with respect to numerical values ​​is arbitrary and may mean, for example, + / - 10%.

[0056] 1 He,Kaiming;Zhang,Xiangyu;Ren,Shaoqing;Sun,Jian(2015-12-10).''Deep Residual Learning for Image Recognition''.arXiv:1512.03385 2 He,Kaiming;Zhang,Xiangyu;Ren,Shaoqing;Sun,Jian(2016).''Deep Residual Learning for Image Recognition''.Proc.Computer Vision and Pattern Recognition(CVPR),IEEE. 3Krizhevsky,Alex,Ilya Sutskever,and Geoffrey E.Hinton.‘‘Imagenet classification with deep convolutional neural networks.’’Advances in neural information processing systems25(2012):1097-1105. 4 Szegedy et al.(2014)‘‘Going Deeper with Convolutions’’arXiv:1409.4842 5 Simonyan&Zisserman(2015)‘‘Very Deep Convolutoinal Networks for Large-Scale Image Recognition’’arXiv:1409.1556 6 Tan&Le(2019)‘‘EfficientNet:Rethinking Model Scaling for Convolutional Neural Networks’’arXiv:1905.11946 7 Maximilian Ilse,Jakub M.Tomcak,Max Welling(2018)‘‘Attention-based Deep Multiple Instance Learning’’.arXiv:1802.04712 8 ‘‘Adam:A Method for Stochastic Optimization’’by Kingma et al https: / / arxiv.org / abs / 1412.6980 9 ‘‘A disciplined approach to neural network hyper-parameters:Part1--learning rate,batch size,momentum,and weight decay’’by Smith https: / / arxiv.org / abs / 1803.09820 10 ‘‘fastai:A Layered API for Deep Learning’’by Jeremy Howard,Sylvain Gugger https: / / arxiv.org / abs / 2002.04688 11 ‘‘PyTorch:An Imperative Style,High-Performance Deep Learning Library’’by Paszke et al,https: / / arxiv.org / abs / 1912.01703

Claims

1. 1. A computer-implemented method for distinguishing between lymphoblast and myeloblast cells, comprising: receiving a digital image including one or more blast cells; applying image analysis to the digital image, the image analysis algorithm comprising: detecting one or more blast cells in the digital image; generating a bounding box around each of the one or more blast cells; generating a plurality of image files, each image file containing a digital image of a blast cell of the one or more blast cells, the boundary of the image in each file corresponding to a respective bounding box; applying image analysis, applying a deep neural network classifier to each of the generated image files, the deep neural network classifier configured to generate an output indicative of whether each blast cell is a lymphoblast cell or a myeloblast cell, the output comprising a number x in the range [0, 1]; If the value x is equal to 1, the blast cells are identified as lymphoblast cells with 100% confidence, and if the value x is equal to 0, the blast cells are identified as myeloblast cells with 100% confidence; or A value of x equal to 1 identifies the blast cells as myeloblast cells with 100% confidence, and a value of x equal to 0 identifies the blast cells as lymphoblast cells with 100% confidence. applying a deep neural network classifier; calculating a patient-level score based on each output value x for all of the blast cells in the digital image, wherein the patient-level score comprises a mean, a median, a maximum value of x, and / or a minimum value of x; 10. A computer-implemented method comprising:

2. The deep neural network classifier is a Resnet34 classifier, a Resnet50 classifier, or a Resnet101 classifier. The computer-implemented method of claim 1 .

3. the digital image is a bright field microscope image containing one or more Romanowsky stained blast cells; 3. The computer-implemented method of claim 1 or 2.

4. and generating instructions based on an output of the parametric model classifier, the instructions configured to cause a display device of the computing system to display a gallery, the gallery comprising: a first plurality of images showing blast cells identified with the highest confidence as lymphoblast cells; a second plurality of images showing blast cells identified with the highest confidence as myeloblast cells; 3. The computer-implemented method of claim 1, comprising:

5. The deep neural network classifier receiving a plurality of pairs of labeled training data, each pair of labeled training data comprising: input data including digital images of blast cells from patients diagnosed with either acute myeloid leukemia or acute lymphoblastic leukemia; output data including an indication of whether the patient has acute myeloid leukemia or acute lymphocytic leukemia, the output including a number x in the range [0, 1]; Including, If the value x is equal to 1, the blast cells are identified as lymphoblast cells with 100% confidence, and if the value x is equal to 0, the blast cells are identified as myeloblast cells with 100% confidence; or If the value x is equal to 1, the blast cells are identified as myeloblast cells with 100% confidence, and if the value x is equal to 0, the blast cells are identified as lymphoblast cells with 100% confidence. receiving a plurality of pairs of labeled training data; training the deep neural network classifier using the training data; 3. The computer-implemented method of claim 1 or 2, wherein the computer-implemented method is generated using a computer-implemented method comprising:

6. 3. A computing device having a processor configured to perform the computer-implemented method of claim 1 or 2 to provide a provisional diagnosis of acute myeloid leukemia or acute lymphocytic leukemia, The processor: determining whether the patient whose blast cells are represented in the digital image has acute myeloid leukemia, acute lymphocytic leukemia, or neither leukemia, based on a patient-level score calculated based on the output of the deep neural network classifier; generating, based on a result of the determining, instructions configured to cause a display device of the computing system to display the result of the determining; a clinical decision support system further comprising:

Images