CROSS-REFERENCE TO RELATED APPLICATIONS

[0001] This application claims the benefit of U.S. provisional application Serial No.

62/815,736 filed March 8, 2019, the disclosure of which is hereby incorporated in its entirety by reference herein.

TECHNICAL FIELD

[0002] In at least one aspect, the present invention is related to machining learning methods for histopathological and clinical outcome based on histopathological feature classification of tissue samples.

BACKGROUND

[0003] Although deep learning (DL) has the potential to teach us novel aspects of biology, the most impressive use cases to date recapitulate patterns that experts already recognize. ^{1 6} While these approaches may improve inter-observer variability and accelerate clinical workflows, our goal is to use DL to learn how morphology from hematoxylin and eosin (H&E) images can be used to predict biomarkeis, ⁷ prognosis* and theragnosis— tasks which are not currently possible for a pathologist to do by eye, but if possible, could improve our understanding of cancer biology. However, since DL learns from data, it needs large training sets to learn patterns. One of the biggest challenges is obtaining large, well-annotated training sets.

[0004] While typical computer vision datasets contain on the order of millions of annotated images to learn statistically significant relationships, ¹ clinical pathology ease sets generally number in the hundreds. Moreover, noise in the clinical annotations dilutes the learning signal and increases the probability the network will leant spurious features like stain color, clinical site, or other technical variations. ^{9 11} [0005] Accordingly, there is a need for improved methods for implementing machine-learned histopathologic classification.

SUMMARY

[0006] In at least one aspect, a machining learning technique in which pre-training allows neural networks to learn histology-optimized features is provided. The method includes a step of training an untrained machine learning device executing a neural network to form a trained machine learning device. The training is accomplished by receiving a digital image of a histologic sample as input to a tissue fingerprinting function and outputting a vector of numbers called a tissue fingerprint that describes histologic patterns within the digital image and training the untrained machine learning device with digital images from a plurality of characterized or uncharacterized stained tissues samples. Characteristically, the untrained machine learning device learning the tissue fingerprinting function through an optimization procedure that minimizes an objective loss function that includes a first component promoting learning of the tissue fingerprinting function that can be used to match digital image patches from the same sample but distinguish digital image patches from multiple samples and an optional second component of the objective loss function promoting learning of the tissue fingerprinting function that is invariant to sources of pathology artifacts that arc encountered during tissue processing. The method also includes a step of predicting a status of a stained tissue sample of unknown status by obtaining a sample digital image for a stained tissue sample of unknown status, applying the trained machine learning device to the sample digital image to determine a predicted tissue fingerprint for the stained tissue sample of unknown status, and finding a best match of the predicted tissue fingerprint to a set of tissue fingerprints for samples of known status.

[0007] In still another aspect, clinical estrogen receptor (ER), progesterone receptor (PR), and Hcr2 status in breast cancer were evaluated with the methods set forth herein. These biomarkers arc important predictive and prognostic molecular markers currently assessed by molecular Immunohistochcmistry (1HC) staining. ¹⁴ BRIEF DESCRIPTION OF THE DRAWINGS

[0008] FIGURE 1 A. Networks are first trained to learn tissue fingerprints, which are patterns of cells and tissue visible on H&E images that can be used to distinguish between patients. Following this training internship, which can be sealed to very large numbers of patients without clinical outcome annotations, the fingerprints are repurposed to make clinically relevant predictions from small labeled datasets.

[0009] FIGURE IB. Schematic of a histopathology classification method using tissue fingerprints.

[0010] FIGURE 1C. Schematic of a convolutional neural network used in the method depicted in Figures lA and IB.

[0011] FIGURE 2. Schematic of a computing device for implementing the machine learning method set forth herein.

[0012] FIGURE 3. Table 1 providing Fingerprinting Datasets.

[0013] FIGURE 4. CycleGAN normalizes the staining colors and styles of tissue images while preserving morphology. Top: the style of slide 1 is translated into the style of slide 2. Bottom: the style of slide 2 is translated into the style of slide 1.

[0014] FIGURES 5A, 5B, and SC. A: Representative tSNE visualization of fingerprints from the test set. In this visualization, left halves from slide 5 and right halves of slide 4. B: Visualization of a representative pair. Left half presented on top, right half on the bottom, the middle shows a heat map of fingerprint distance (distance from fingerprints from the bottom image to the average fingerprint of the top image). C: Left, exploded displays of the original patches in the embedding show similar histologic features (nucleoli, micro-papillac, fat, mucin).

[001S1 FIGURES 6A, 6B, and 6C. A: heat maps of areas that lead to accurate patient classification. Higher probability regions (red) are more predictive of patient identity, and hence distinctive, than blue regions. B: an exploded view of two cores from (A). C: Legend [0016] FIGURE 7A. Illustration of whole slide clinical ER classification. An analogous procedure was used for PR and Hcr2 classification. Fingerprints were extracted from 120 random image patches, and a second ER-classificr, acting on the fingerprints, made local predictions, which were averaged to produce a continuous whole-slide-level ER-score.

[0017] FIGURES 7B-1, 7B-2, and 7B-3. B: Receiver operating characteristic curves (ROC) for clinical ER (left), PR (center), and Her2 prediction (right). The TCGA ROC curve reflects a test set from five-fold cross-validation, and the AUC corresponds to the average area under the ROC curves of all five TCGA test sets. All samples in the ABCTB dataset are test samples and were never seen during training. Sample sizes vary depending on the availability of clinical annotations

[0018] FIGURES 8A, 8B, and 8C. a: Histogram of ER-predictions from the TCGA test set averaged across the entire slide (AUC = 0.88). B: AUC scores obtained by pooling ER predictions from different regions within slides. C: Representative heatmaps of correctly and incorrectly classified whole slides. WS1 prediction was obtained by averaging over all patches (epithelium, stroma, fat). Each slide visualization consists of an RGB thumbnail, a tissue type segmentation, and an ER prediction hcatmap.

[0019] FIGURE 9. Left: tSNE embedding of fingerprints from patches extracted from TCGA whole slides, shaded by ER prediction score. 12 clusters with high positive or negative enrichment were selected for manual inspection. Right: H&E patches closest to the cluster centers. Each patch is from a different patient

DETAILED DESCRIPTION

[0020] Reference will now be made in detail to presently preferred compositions, embodiments and methods of the present invention, which constitute the best modes of practicing the invention presently known to the inventors. The Figures are not necessarily to scale. However, it is to be understood that the disclosed embodiments arc merely exemplary of the invention that may be embodied ui various and alternative forms. Therefore, specific details disclosed herein are not to be interpreted as limiting, but merely as a representative basis for any aspect of the invention and or as a representative basis for teaching one skilled in the art to variously employ the present invention.

[0021] It is also to be understood that this invention is not limited to the specific embodiments and methods described below, as specific components and or conditions may, of course, vary.

Furthermore, the terminology used herein is used only for the purpose of describing particular embodiments of the present invention and is not intended to be limiting in any way.

[0022] It must also be noted that, as used in the specification and the appended claims, the singular form "a, " " an," and "the" comprise plural referents unless the context clearly indicates otherwise. For example, reference to a component in the singular is intended to comprise a plurality of components.

[0023] As used herein, the term“about" means that the amount or value in question may be the specific value designated or some other value in its neighborhood. Generally, the term“about” denoting a certain value is intended to denote a range within + -5° · of the value. As one example, the phrase“about 100” denotes a range of 100-. -5, i.e. the range from 95 to 105. Generally, when the term“about” is used, it can be expected that similar results or effects according to the invention can be obtained within a range of +'-5°e of the indicated value.

[0024] The term“and or" means that either all or only one of the elements of said group may he present.

[0025] It is also to be understood that this invention is not limited to the specific embodiments and methods described below, as specific components and or conditions may. of course, vary. Furthermore, the terminology used herein is used only for the purpose of describing particular embodiments of the present invention and is not intended to be limiting in any way.

[0026] It must also be noted that, as used in the specification and the appended claims, the singular form "a," "an." and "the" comprise plural referents unless the context clearly indicates otherwise. For example, reference to a component in the singular is intended to comprise a plurality of components. [0027] The term“comprising" is synonymous with“including,"“having,"“containing," or

“characterized by." These terms arc inclusive and open-ended and do not exclude additional, unrccitcd elements or method steps.

[0028] The phrase“consisting of excludes any element, step, or ingredient not specified in the claim. When this phrase appears in a clause of the body of a claim, rather than immediately following the preamble, it limits only the element set forth in that clause: other elements arc not excluded from the claim as a whole.

[0029] The phrase“consisting essentially of limits the scope of a claim to the specified materials or steps, plus those that do not materially affect the basic and novel charactcristic(s) of the claimed subject matter.

[0030] The phrase“composed of ^* means“including" or“consisting of. Typically, this phrase is used to denote that an object is formed from a material.

[0031] With respect to the terms“comprising,"“consisting of." and“consisting essentially of," where one of these three terms is used herein, the presently disclosed and claimed subject matter can include the use of either of the other two terms.

[0032] The term“one or more" means“at least one" and the term“at least one" means“one or more." The terms“one or more" and“at least one" include“plurality" as a subset.

[0033] The term“substantially,"“generally." or“about" may be used herein to describe disclosed or claimed embodiments. The term “substantially" may modify a value or relative characteristic disclosed or claimed in the present disclosure. In such instances,“substantially" may signify that the value or relative characteristic it modifies is within =. 0.1%·, 0.5%, 1 ·. 2%, 3%, 4% 5% or 10% of the value or relative characteristic.

[0034] It should also be appreciated that integer ranges explicitly include all intervening integers. For example, the integer range 1-10 explicitly includes 1 +- 4, 5, 6, 7, 8, 9, and 10. Similarly, the range 1 to 100 includes 1, 2, 3. 4. . . . 97.98, 99, 100. Similarly, when any range is called for, intervening numbers that arc increments of the difference between the upper limit and the lower limit divided by 10 can be taken as alternative upper or lower limits. For example, if the range is I . I . to 2.1 the following numbers 1.2, 1.3, 1.4, 1.5, 1.6, 1.7, 1.8. 1.9, and 2.0 can be selected as lower or upper limits. In the specific examples set forth herein, concentrations, temperature, and reaction conditions (c.g. pressure, pH, etc.) can be practiced with plus or minus 50 percent of the values indicated rounded to three significant figures. In a refinement, concentrations, temperature, and reaction conditions (c.g., pressure. pH, etc.) can be practiced with plus or minus 30 percent of the values indicated rounded to three significant figures of the value provided in the examples. In another refinement, concentrations, temperature, and reaction conditions (e.g., pH, etc.) can be practiced with plus or minus 10 percent of the values indicated rounded to three significant figures of the value provided in the examples.

[0035] The term“computing dev ice ^** refers generally to any device that can perform at least one function, including communicating with another computing device. Sometimes the computing device is referred to as a computer.

[0036] When a computing device is described as performing an action or method step, it is understood that the computing devices arc operable to perform the action or method step typically by executing one or more lines of source code. The actions or method steps can be encoded onto non- transitory memory (e.g., hard drives, optical drive, flash drives, and the like).

[0037] The term“neural network" refers to a machine learning model that can be trained with training input to approximate unknown functions. In a refinement, neural networks include a model of interconnected digital neurons that communicate and leant to approximate complex functions and generate outputs based on a plurality of inputs provided to the model.

[0038] The term“subject ^** or“patient ^** refers to a human or other animals, including birds and fish as well as all mammals such as primates (particularly higher primates), horses, birds, fish sheep, dogs, rodents, guinea pigs, pig, cat, rabbits, and cows.

)0039) The term“biomarker” refers to any biological properly, biochemical feature, or aspect that can be used to determine the presence or absence and'or the severity of a disease or disorder such as cancer. [0040] Throughout this application, where publications are referenced, the disclosures of these publications in their entireties arc hereby incorporated by reference into this application to more fully describe the state of the art to which this invention pertains.

[0041] Abbreviations:

[0042] “ABCTB" means the Australian Breast Cancer Tissue Bank.

[0043] “AUC" means the area under the curve.

[0044] ”ER” means estrogen receptor.

[0045] “H&E" means Hematoxylin and cosin.

[0046] “HER2” means human epidermal growth factor rcccptor-2.

[0047] "NN” means neural network.

[0048] “PR“ means progesterone receptor.

[0049] “USC” means the University of Southern California.

[0050] “TCGA" means The Cancer Genome Atlas.

[0051] "TMA ^** means tissue microarray.

[0052] "IHC" means immunohistochcmistry.

[0053] In general, the present invention provides a machining learning method in which pre- training on histology images allows neural networks to learn histology-optimized features that would improve performance on subsequent clinical classification tasks. In achieving this goal, the concept of“tissue fingerprints' ^* is introduced. Tissue fingerprints arc based on the hypothesis that molecular differences of the tumor arc oficn translated into subtle differences in morphologic phenotypes. This idea is akin to the paradigm of precision medicine, w here instead of grouping patients, individual patients are treated based on their specific molecular and environmental parameters. Hence, instead of training a network to distinguish between groups of samples, we first pre-configure the network to recognize, or“fingerprint," ""individual tumors, a task which can leverage large, unannotated datasets, which are widely available. After pretraining a network to fingerprint tissues, we expect that a smaller amount of annotated data will be necessary to adapt it to a clinical task.

[0054] As set forth below in more detail, tissue fingerprints are implemented by training a neutral network to fingerprint pathologic tumor samples from a training set and then tested it on a simple matching task using tumor images from new patients. Briefly, multiple images were divided into halves, and a network attempted to learn a vector of features (a“fingerprint”) that could correctly pair the halves (Figure 1 ). An important aspect of this work was using the matching task to lcam stain- and site- invariant features of the architecture. We controlled for these sources of noise by testing whether the fingerprints could match tissues from the same patients that had been stained and scanned at different sites. Optimizing on the task of matching, we performed experiments testing the impact of training set size and methods of image normalization. The features (fingerprints) were extracted from the prc-wired fingerprint network after training internship was accomplished and used to classify between groups of tumors with biologically relevant molecular pathway annotations. ^12.13

[0055] With reference to Figures 1A, 1B, and 1C, a histopathology classification method through machine self-learning of "tissue fingerprints” is illustrated. Characteristically, the histopathology classification method is implemented in a digital medium environment. Figure 1A shows that in general, the method includes a first step of training a neural network to lcam the tissue fingerprints and a second step of applying the fingerprint to a clinical task. Figure 1 B provides a more detailed schematic of the method. With reference to Figures 1 A and 1 B, histopathology classification method 10 includes a step of training an untrained machine learning device 12 executing a neural network (i.e.. computer instructions for a neural network) to form a trained machine learning device 14. The training step is accomplished by receiving a digital image 16 ¹ of a histologic sample as input to a tissue fingerprinting function 18 and outputting a vector of numbers called a tissue fingerprint (TFP) 20 that describes histologic patterns within the digital image. The untrained machine learning device 12 is trained with digital images 16 ⁿ from a plural ity of characterized or uncharactcrizcd stained tissue samples 20 ^p where n is an integer equal to the number of digital images and p is the number of tissue samples. The number n of digital images and the number p of characterized or uncharactcrizcd stained tissues samples 20 ^p can independently be from 10 to several hundred or more. The untrained machine learning device 12 lca5rns the tissue fingciprinting function 16 through an optimization procedure 22 that minimizes an objective loss function 24. Details of the object loss functions arc set forth below in the experimental section. The objective loss function 24 includes a first component 26 promoting the learning of the tissue fingerprinting function that can be used to match digital image patches 30 from the same sample but distinguish digital image patches 32 from multiple samples. The objective loss function 18 also includes an optional second component 36 that promotes learning of the tissue fingerprinting function 18 that is invariant to sources of pathology artifacts that arc encountered during tissue processing. Examples of such artifacts include, but arc not limited to. pathology artifacts that include variation in tissue stain concentrations, in tissue processing procedures, differences in slide scanning equipment, and combinations thereof. In a refinement, multiple tissue fingerprints across different tissue regions arc used in a pooled fashion to determine tissue status. It should also be appreciated that samples arc derived from (i.c.. obtained from) sub-regions of tissue within whole slide images from either a tissue blocks or core biopsy specimens. In a refinement, the objection loss function further includes a third optional component 38lo apply an image normalization step before training.

[0056] The method also includes a step 2 of predicting the status of a stained tissue sample 40 of unknown status. This step is accomplished by obtaining a sample digital image 42 for a stained tissue sample of unknown status 40. The trained machine learning device 14 is applied to the sample digital image to determine a predicted tissue fingerprint 44 for the stained tissue sample 40 of unknown status. Finally, a best match 46 of the predicted tissue fingerprint to a set 48 of tissue fingerprints for samples of known status is found.

[0057] In another refinement, a relationship between a collection of characterized samples and their fingciprinls is learned and then applied to predict a status for the stained tissue sample of unknown status.

[0058] In a variation, learning of a tissue fingerprinting function 18 that is invariant is promoted by obtaining paired images 50 and penalizing the optimization procedure for learning a fingerprint function that assigns different fingerprints to the paired images. Typically, the paired images 50 can be obtained either by sectioning a tissue sample that is processed at different laboratories, or by altering pixels of an original image to give an appearance of having been processed at a different laboratory.

[0059] In another variation, a relationship is learned between the tissue fingerprint and a diagnostic, prognostic, or Iheragnoslic data annotation of the sample. From these features, a therapeutic treatment strategy can be determined for a subject from a subject's predicted status for the status of a diagnostic feature w ith which the subject is subsequently treated. In a refinement, the subject is treated with a cancer therapeutic agent.

[0060] In another variation, the diagnostic, prognostic, or thcragnostic feature, is the presence or absence of a biomarker. Typically, in this variation, the stained tissue sample of unknown status is a cancer sample. Such a cancer sample may be a cancer of unknown origin, such as cancer from a metastatic site. In a refinement, the tissue fingerprinting function is used to infer the site from which the cancer originated (tissue of origin of the cancer). Examples of useful biomarkers include, but arc not limited to. estrogen receptor (HR), human epidermal growth factor receptor-2 (HHR2), progesterone receptor (PR), proliferation marker Ki-67, and cytokcratin markers. When the biomarkcr is ER, a predicted status can be used to determine specific treatments. In a refinement, the biomarkcr is FR, PR. and HER2 with a predicted status indicating prognosis.

[0061] As set forth above, the untrained machine learning device 12 (and therefore, untrained machine learning device 1 ) is a computing device (c.g., a computer) executing instructions for the neural network. Typically, the neural network is a convolutional neural network. In this regard, the convolutional neural network includes a plurality of convolutional layers and a plurality of pooling layers. Figure 1C provides an idealized schematic illustration of a convolutional neural network executed by untrained machine learning device 12. Typically, the neural network is a convolutional neural network. In this regard, the convolutional neural. It should be appreciated that any deep convolutional neural network that operates on the pre-processed input can be utilized. The convolutional network can include convolutional layers, pooling layers, fully connected layers, normalization layers, a global mean layer, and a batch-normal ization layer. Convolutional neural network layers can be characterized by sparse connectivity where each node in a convolutional layer receives inpul from only a subset of the nodes in the next lowest neural network layer. The Convolutional neural network layers can have nodes that may or may not share weights with other nodes. In contrast, nodes in fully-connected layers receive input front each node in the next lowest neural network layer. For both convolutional layer and fully connected layers, each node calculated its output activation front its inputs, weight, and an optional bias. During training, optimal values for the weight and bias arc determined. For example, convolutional neural network 60 receives a plurality of digital images 16" which arc processed as set forth above. Convolutional neural network 60 includes convolution layers 62. 64. 66, 68. 70, and 72 as well as pooling layers 74, 76. 78, 80 and 82. optional processing layer! s) 84. and multiple outputs 88 ^a (where n is as above). The pooling layers can be max pooling layer or a mean pooling layer. Each of the convolution layer can include multiple internal feature maps w hich can change front layer to layer via subsampling and convolution. The present embodiment is not limited to by the number of convolutional layers, pooling layers, fully connected layers, normalization layers, and sublayers therein.

[0062] As set forth above, the histopathology classification methods are implemented by a computing device. In general, computing devices arc computer processor-based electronic devices. With reference to Figure 2, computing device 100 includes computer processor 102 that executes the instructions for authoring the chemical mechanism problem or solving it. It should be appreciated that virtually any type of computer processor may be used, including microprocessors, multi-core processors, and the like. The instructions for the method typically arc stored in computer memory 104 and accessed by computer processor 102 via connection system 106. In a v ariation, connection system 106 is and or includes a data bus. In a refinement, computer memory 104 includes a computer- readable medium 108 which can be any non-transitory (c. g., tangible) medium that participates in providing data that may be read by a computer. Specific examples for computer memory 104 include, but arc not limited to, random access memory (RAM), read-only memory (ROM), hard drives, optical drives, removable media (c.g.. compact disks (CDs), DVD, Hash drives, memory cards, etc.), and the like, and combinations thereof. In another refinement, computer processor 102 receives instructions from computer memory 104 and executes these instructions, thereby performing one or more processes, including one or more of the processes described herein. Computer-executable instructions may be compiled or interpreted from computer programs created using a variety of programming languages and or technologies including, without limitation, and cither alone or in combination, Java, C, C l I , C#, Fortran. Pascal, Visual Basic, Java Script. Perl, PL· SQL, etc. Display 1 10 is also in communication with computer processor 102 via connection system 176. Computer device 170 also includes various in out ports 1 12 through which data from a pointing device 1 14 may be accessed by computer processor 102. Examples for the computing device include, but are not limited to, desktop computers, laptops, or servers. Examples of pointing devices include a mouse, touch screen, stylus, trackball, joystick or touchpad. In a particularly useful variation, the pointing device is incorporated into display 108 as a touch screen by which user 1 16 interacts with a linger. In a variation, a non- iransilory storage medium or media (hard drives, optical drives, removable media (c.g.. compact disks (CDs), DVD, flash drives, memory cards, etc.) has encoded thereon instructions for the steps executed by computer processor 102 in performing the steps of the methods set forth herein.

[0063] The following examples illustrate the various embodiments of the present invention. Those skilled in the art will recognize many variations that arc within the spirit of the present invention and scope of the claims. It should be appreciated that the techniques set forth below can analogously be applied to the other biomarkers set forth above.

[0064] METHODS

[0065] Tissue Fingerprinting

[0066] Dataset

]0067] The tissue fingerprint network was trained on images of tissue microarray (TMA) cores. The tissue microarrays (TMAs) used in this study were obtained from supplier US Biomax. Inc. Array BR20823, containing 207 tissue cores from 104 patients, was used to train the fingerprint network. Array BR20819. containing 208 cores from a separate group of 104 patients, was used to test the trained model. One or two sections were obtained from each array (for BR20823 and BR20819, respectively), which were H&E stained using standard laboratory protocols, before scanning was performed at 40x resolution (0.249 microns per pixel) on a Carl Zeiss slide scanner. US Biomax, Inc. kindly provided images of serial sections of these microarrays that had been stained and scanned by their protocols on an Apcrio slide scanner at lower resolution (0.49 microns per pixel). Additional breast cancer tissue images were used to increase the size of the training set for experiments 2 and 4. These images arc of breast cancer tissue from a variety of sources having distinct patients from BR20819 and BR20823. (sec. Table I in Figure 3)

[0068] Neural Network Training (experiments 1 and 2)

[0069] Neural networks were trained and run on NVIDIA PI 00 GPUs (Oracle Cloud Infrastructure BM.UPL ^' 2.2 instances). The fingerprint network was trained on image patches randomly extracted from the BR20823 images. Each circular tissue core was isolated, scaled to 0.5 micron-pixel resolution (bilinear resampling), and cropped to a 1600.x 1600 pixel square. These squares were assigned a numeric index from 1 to 207, reflecting their position on the array. Each square was then divided into left and right halves. During training, a small image patch (224 x 224 px) was sampled from the left half, augmented through rotation, color spectrum augmentation ¹ , color normalization. It was then passed to a neural network trained to minimize cross entropy loss (i.c., an example of an objective loss function) between patient identity and a predicted index. During training, progress was monitored by measuring how well the network could predict the con; index from patches from the right halves of the tissue images, which it hadn’t seen. When this accuracy platcaucd, training was stopped and the quality of the features on the tissue matching game was tested. Experiments 3 and 4 used the more complex objective loss function described below.

[0070] In the four experiments described, the standard implementation of the Rcsnct34 architecture ¹⁵ provided by the PyTorch library ¹ was used. The network was randomly initialized and trained from scratch. Additionally, larger networks, including Resnet50 and Resnet 100 . were trained using the conditions of experiment 4, but found the same performance as Rcsnet34 (results not shown). The benefit of larger networks may depend on the size of the training dataset. Here the concept of fingerprinting with a relatively small dataset is demonstrated, but training on larger datasets may demonstrate added benefits of deeper networks.

[0071] Promoting style Invariance by GAN-based style transfer (experiments 3 and 4):

[0072] Neural style transfer was performed offline using CyclcGAN ^1" , a method of neural style transfer, that aims to alter the style of images while preserving fine details. Briefly, the CyclcG AN approach consists of training two neural nets, a generator which lakes an image A transforms it into an image of style B. and a discriminator which is trained to distinguish between generated images and real ones. The networks arc trained simultaneously as adversaries. As the discriminator improves, the generator is challenged to learn better transformations from style A to B. Conversely, as the generator improves, the discriminator is challenged to learn better features that distinguish real and generated images. In this project, the open-source CyclctiAN code was used without modification. The network was trained to transfer styles between images of BR20823 that were stained by the array manufacturer (US Biomax) or at our site (style transfer between slides 1 and 2. respectively, as shown in Figure 4). Thus, the original set of 13.415 cores w as augmented to threefold its original size via neural style transfer (each core has an original image, a virtual USC stain and a virtual Biomax stain). Following style transfer, the objective loss function to promote style invariance was adapted. The new objective loss function has two components, a cross entropy loss (abbreviated‘CE ^* ) lo predict the identity of each patch, which was the loss term used in experiments I and 2, plus an additional loss term to minimize the distance of fingerprints from different styles. The additional loss term is the squared error (abbreviated‘SE') between the L2-normalizcd fingerprints, where a fingerprint is the 512-dimensional (512D) feature vector from the last layer of the Resnet.

[0073] Loss is defined for a pair of images front tissue core, ( I £ i £ 207): Image, i, and Image,:.

Image·: is a re-styled version of Image, i that contains the same morphological information. The loss is a sum of two cross entropy losses and a fingerprint distance loss. The symbol, g, is a constant. In our experiments, we used g = 0.5.

[0074] Cross entropy is defined using the classification vector produced from the network for each image. This vector contains 207 elements. Image, _» from core, produces a classification vector c, _» .

[0076] Fingerprint distance loss is defined using the fingerprints produced by the neural network for each of the images. If the fingerprints for Image, i, and Image,: arc called FP,i and FPi:, fingerprint distance loss is the following. || || and SE refer to the 1.2 norm and the Euclidean distance, respectively.

[0078] Creating heat maps of recognized regions

[0079] Heat maps of tissue cores showing the parts of an image most predictive of tumor identity (Figure 6) were generated. The region colors were determined as follows: each core image was divided into overlapping square patches (size 224x224 pixels, with 80% linear overlap). Each patch was passed to the neural network, and a probability vector was calculated predicting the identity of each core via Softmax. Since there arc multiple cores per patient, the 207 probabilities were aggregated into 104 probabilities (one per patient) by summing the probabilities of cores that came from the same patient. Each heat map shows the probability of predicting the correct patient and is shaded from 0 (blue-light) to I (red-dark).

[0080] Stain normalization is necessary for fingerprinting

[0081] Four experiments were performed, varying training set size and image normalization, to determine how to best train a fingerprint network (Table 2). It is hypothesized that training on large numbers of cores would improve accuracy, and that image color normalization would greatly improve training on small data but have a smaller effect on larger datasets. In experiment 1 , the 207 tissue cores were collected from slides I and 2 (serial sections of the training TMA that were stained by the supplier and USC, respectively), divided them in half, and trained the network to recognize patients based on patterns in one of the halves. (It was arbitrarily chosen to train on the left halves). Each core was assigned a number from 1 to 207, and the network was trained to identify the number from a patch sampled from the image-half. In experiment 2. the dataset was scaled over 20-fold: adding 13.000 additional training images. Again, the network was trained to predict the index. In experiments 3 and 4. the same datasets were used as before, but included a color normalization procedure based on neural style transfer. I ^* .20 In the first two experiments, it was predicted that as the network was trained to recognize increasing numbers of images, it would automatically learn stain-invariant features. In the second two experiments, the style transfer algorithm CyclcGAN ¹⁷ was used to recolor images (Figure 4). making them appear as if they were prepared at a different site. CyclcGAN can exchange the texture between two spatially similar image sets, while preserving overall structural information. Compelling examples include transforming photographs into impressionist paintings and horses into zebras. Here, wc use CyclcGAN to transfer the H&E staining coloration from a reference site to images from other sites. Then wc trained the networks to look at both images and predict the same features, cancelling out the effect of stain variation.

[0083] ER, PR, Her2 Classification from whole slides

[0084] Datasets

)0085) The whole slide images used in this study were obtained from The Cancer Genome Atlas ¹² (TCGA) and the Australian Breast Cancer Tissue Bank ¹ ’ (ABCTB). Wc included 939 cases Breast Carcinoma from TCGA and 2531 Breast Cancer cases front the ABCTB. Clinical characteristics arc summarized in Table 3. Of note, molecular status of ER. PR. and Hcr2 was assessed clinically, where annotations were made on the patient level.

[0086] Table 3. Clinical characteristics of breast cancer patients

[0087] Training a patch-based ER,PR,Her2 classifiers directly from images (control)

[0088] Similar to other deep learning approaches, a patch-based classifier was trained to predict molecular marker status. All experiments were conducted using five-fold cross validation. First, each whole slide image was grossly segmented into foreground vs. background. The foreground areas were divided into non-overlapping squares of (1 12 x 1 12 microns), which were scaled to a final patch si/e of 224 x 224 pixels. 120 patches per patient were randomly selected and used for downstream analysis. To train the classifier, the entire set of TCGA patients was split into five groups. For each cross validation fold, three groups were used to train, one group (the "overfilling group”) was used to monitor overfilling and perform early stopping, and the remaining group was used to test the network’s final performance. To the train the network, patches were assigned a binary label per the patient-level annotation, but early stopping was implemented by averaging the predictions of all patches belonging to overfilling group and measuring the patient-level AUC (area under the ROC’ curve) score. These experiments used a standard implementation of Rcsnct34.

[0089] Training ER, PR, Her2 classifiers directly from fingerprints

[0090] Using the same image patches and cross validation splits front the control experiment (described above) and the previously trained fingerprint network, 512D fingerprints w'crc extracted for each image patch and then trained a second‘"biomarker" neural network to predict marker status based on these fingerprints. The second network has the follow ing structure: input. 512x8 linear layer, rectified linear unit nonlinearity. 8x 1 linear layer, hyperbolic tangent nonlinearity. The parameters of this network were trained in batches consistent with multiple instance learning (sec Training the fingerprint-based classifier to predict molecular markers). Similar to the control experiment, multiple patch predictions were pooled lo make patient-level predictions, and the reported ROC curves compare patient-level score and clinical marker status. As a control for the features, the same workflow using image features extracted from a Rcsnci34 network prctniincd on the ImagcNct datasel was implemented. The parameters for this model were obtained from the Torchvision Python library.

[0091] Training the fingerprint-based classifier to predict molecular markers

[0092] The classifier for each molecular marker was trained in the following manner. From the training set. 120 patches were extracted per patient (as set forth above), and fingerprints were extracted using the prc-iraincd fingerprint network. Then, a matrix containing these fingerprints was created and loaded into memory. The dimensions of the matrix w ere (N x 120 x 512). where N corresponds to the number of patients in the training set. The tensor was reshaped into a data matrix "Xtrain.’’ with the follow ing dimensions (N * 120 x 512) Additionally a "ground truth' vector {length - N) containing the ground truth data for the training patients was prepared. Each value in the vector signifies positive or negative biomarkcr status by the value of * 1 or -1 X train was passed to a neural network with the following structure:

Linear (512. 8)

ReI.U

Linear (8. 1)

Tanh

The output of the network after receiving X train is of size (N* 120. I ) and is labeled “Pred_unscaled” The matrix. "Pred_unscaled' is reshaped to size: N x 120 Each row in the matrix is normalized as follows:

o The absolute value is taken of Pred_unscalcd and summed over the horizontal dimension, producing a "weight ¹ vector of length X

o Each row in Pred_unscalcd is divided by the corresponding weight \ color, producing a matrix Pred_scaled of shape (N x 120)

• For each row i (0..N):

• Pred scaled(i) = Pred_unscaled[i] weight[i]

o The output is a matrix of shape N x 120

[0093] Each row in the matrix Prcd_scalcd is summed to produce a 1 D“prediction” vector for the training dataset of length N. The loss is calculated by the binary cross entropy between the prediction vector and the ground truth vector. The weights of the model arc updated using back propagation and gradient decent. We used the Pytorch implementation of the Adam optimizer with learning rate of 0.0001 Training was continued until the loss on the cross-validation set stopped decreasing as per the methods section. The motivation for applying the normalization step using the average of the absolute value was to allow the network to pay attention to certain areas and ignore others. The Tanh nonlinearity outputs a score from - 1 to 1. Areas predicted to be 0 are not contributory to the weight and hence do not factor into the final score for the slide. However, areas that arc strongly positive or negative affect the weighting and influence the final patient prediction. [0094] External validation of the whole-slide ER, PR, or Her2 classifier

[0095] After the five-fold cross validation experiments on the TCGA dataset, the classifier was validated on an independent test set from the ABCTB. These images were processed like those from TCGA: 120 patches ( 1 12 x 1 12 microns, resized to 224 x 224 pixels) were extracted per patient, and fingerprints were calculated using the pre-lrained fingerprint network. The TCGA-lraincd biomarker network was then used to predict marker status of patients in the ABCTB dataset. An AUC score w as calculated from the ROC' curve comparing the neural network predictions to clinical status from the ABCTB.

[0096] Visualization of regions used to predict clinical ER score

[0097] Generating heatmaps of predicted ER status

[0098] Given the high AUC ^’ for the ER classifier, hcalmaps showing the regions predicted to be highly ER-positivc or negative across whole slides were made. Every very' non-overlapping foreground patch ( 1 12 x 1 12 micron) were extracted from TCGA VVSIs and compressed them into a 5121) fingerprints using the pretrained fingerprint network. Subsequently, the pre-trained ER network was used to make patch-level predictions across the entire slide. The predictions arc shaded in grayscale. Black signifies a prediction of -1 (ER-negative), while white signifies -1 (ER-positivc). Gray values correspond to scores dose to 0 (indeterminate). All visualizations of ER predictions and features show the validation and test sets of one cross-validation split. Hence, none of the data reflect images that were used to update the values of the w eights.

[0099] Heatmaps of tissue types

[0100] TCGA whole slides were segmented into background, epithelium, stroma, and fat by training a patch-based classifier on a publicly available dataset ofWSls with tissue type annotations. ¹ * The segmentations w ere subsequently assessed for gross accuracy at 2x resolution, and a small portion of hcatmaps (<5%) were manually corrected.

[0101] Visualizing the tSNE embedding of WSI fingerprints [0102] An efficient implementation of the tSNF. algorithm was used to plot a 2D manifold of the fingerprints. ¹⁹ Each point represents one fingerprint, which corresponds to a single patch from a

WSI. The color of the point represents the ER prediction, which is shaded from blue (-1 , ER negative) to green (0, neutral) to red (+1 , ER positive). Grossly, the presence of regions containing a predominance of blue or red points was appreciated. To assess the visual similarity of patches in these clusters, 12 cluster centers were manually selected and live patches closest to these centers visualized.

[0103] RESULTS

[0104] In summary, the basic aim of the present invention was to develop a biologically meaningful set of H&E histologic features. It is hypothesized that in the process of training a neural network to "match two halves ^* ' of tissue cores, it would lcam a compact, but meaningful representation of tissue architecture. Following training of the fingerprint network, which was influenced by the discovery of the importance of stain-normalization, various methods of predicting molecular marker status from whole slides were compared. It was found that a fingerprint-based approach out-performed traditional transfer-learning and direct patch-based classification. Moreover, fingerprint-based classifiers continued to perform well on an independent, external dataset. When the fingerprint-based classifier was applied to a small collection of slides in the test-set, it was found that they produced interpretable heatmaps, and predominantly focus on epithelial patterns to make predictions.

I0I0SJ Learning fingerprints

[0106] The networks were trained to lcam fingerprints from tissue cores in TMAs. The TMA format makes it easy to process one set of tissues in multiple ways and allowed us to simulate the batch-effects that are commonly encountered at pathology labs. By staining and scanning one section at USC and another stained by the TMA supplier (US Biomax), we obtained paired images with the same architectural features, but different coloration. Our goal was to learn a fingerprint that summarized the architecture but ignored the staining differences.

[0107] We used one TMA to train (BR20823) and another TMA to test (BR20819. Table 1 ). Each TMA contains approximately 208 tissue cores from 104 patients, with no patient overlap between arrays. We used 3 serial sections of the training TMA. One section was stainedscanncd by the TMA supplier (slide I ), the other two were stained at USC (slides 2. 3). Two serial sections of the lest array were similarly collected, stained at USC and by the TM A supplier (slides 4 and 5).

[0108] To compare the quality of the fingerprints learned in the four experiments, tissue matching on the test TMA sections was performed. Using the thus-trained NX. we calculated fingerprints for left halves of cores from one section (stained by the array manufacturer, slide 4) and the right halves from the other (stained at USC, slide 5), and matched each left fingerprint to the nearest right fingerprint in 512D fingerprint space. Since them were 208 cores in the test set, we report a core- level accuracy (acc. = number of cores matched correctly ' 208). The null accuracy by chance is 0.4° · ( 1 208 cores).

[0109] While fingerprints from all four experiments matched cores better than chance, the accuracy was highest in experiment 4. which used a large training set w'ilh stain normalization. The fingerprints from this method matched cores with 63°'· accuracy (131-208 cores). Surprisingly, stain normalization seems to be necessary to get the performance gains of larger training sets. Comparing the results of experiments 2 and 3 to the baseline, increasing training set size in the absence of stain- normalization (experiment 2) provided only a miniscule improvement in matching accuracy over the baseline. However, stain-nonnalization nearly doubled the accuracy. lt ^* s important to note that in all four experiments we used standard image augmentation during training. Immediately before the image was shown to the fingerprint netw'ork, it was randomly adjusted for brightness and contrast and converted to grayscale. Even with these procedures, which w ere intended to make networks invariant to color differences between images, ⁶ doing an additional style normalization step before the augmentation provided a significant improvement.

[0110] When we examined the mistakes from the network in experiment 4. it was noticed that a large portion of them were due to an aspect of the study design, using a test set with 2 cores per patient. Several of the misclassificd cores were incorrect at the core level but correct at the patient level. This is because some tumors arc morphologically homogeneous and have similar patterns across cores. Thus, we also calculated a pooled accuracy, which uses the fingerprints from both left cores to match both right cores and found that fingerprints could match patients with 93°· accuracy (sec methods for details). [0111] Encouraged by the high patient- level accuracy of the matching, the features of the final neural network layer was studied. When the network is shown an image, this layer produces a 512D vector of real numbers, the tissue fingerprint. In the remainder of the work, the netw'ork w as applied to extract fingerprints and explore how they can be used to link histologic patterns to clinical subgroups of breast cancer.

[0112] Fingerprint visualizations reveal style-invariant histologic patterns

[0113] As both a control and as a starting point for understanding the features in the fingerprints, we performed two-dimensional fingerprint visualization. The left and right halves of cores were taken from the test slides, which had been stained at different sites, and calculated fingerprints for patches from the halves. Next, the fingerprints were embedded in a tSNF plot (Figure 5A). tSNE is a technique that compresses high dimensional vectors into a 2D space while preserving local structures in the data. ²¹ Its primary utility in this work is to approximate the relationships within a high dimensional space and make them accessible for visual analysis. After using tSNE to compress the fingerprints into 21) coordinates and plotting each coordinate as a point colored by patient index, it was observed that the left and right halves from the same patient arc close in the embedding space. Since points that are nearby on a tSNE plot also lend to be nearby in the original space, this visualization provides evidence that fingerprints of similar tissues arc similar in spite of differences in staining. Moreover, visualizing the same embedding as a map of image patches, instead of colored points, reveals that different regions of the map contain different architectures, such as nucleoli, fat, micropapillary growth, and mucin patterns (Figure 50. Thus, in the embedding space, fingerprints are clustered by histologic patterns even if the patches they come from exhibit markedly different colorations and styles.

[0114] Fingerprints can be used to visualize similar regions between tissues

[0115] Figure SB focuses on a specific Lcfl Right pair from the test set. The average fingerprint of the right half is calculated and a heat map plotted show ing the similarity (defined as 1 - normalized Euclidean distance) from each patch in the left half to the average fingerprint of the right half (red is similar, blue is dissimilar). The overlay (bottom) shows that similarity between the right and left halves is highest in a discrete region that appears to contain epithelial cells. This observation is consistent with the abundance of epithelial cells in the right image, suggesting that fingerprint similarity may have utility in histologic search.

[0116] Fingerprints combine epithelium and stromal features

[0117] To directly visualize the image components comprising its fingerprint, heat maps of tissue cores, highlighting image regions that most accurately predict patient identity were generated

(Figure 6A). Figure 6A shows the original H&E images alongside heal maps of patient prediction accuracy using corresponding image regions. Red areas identify the patient accurately, and blue ones do so poorly. Based on the presence of both red and blue areas, some core regions arc more predictive of patient identity than others, meaning their patterns are specific to that patient’s tumor. Figure 6B shows an exploded view of two cores. The red-colored regions demonstrate the classifier looks at a combination of stromal and epithelial areas.

[0118] Fingerprints relate to molecular status of breast cancer

[0119] Because each tumor has unique genetic and microenvironmental interactions, it is hypothesized that networks trained to recognize patients would implicitly leant features that reflect the underlying biological processes. For breast cancer, ER'PR Her2 status is among the most important indicators of prognosis. Hormone-receptor (ER.PR) positive tumors lend to be less aggressive and occur in older patients. Additionally, ER-positive tumors can be treated effectively with drugs that target the estrogen axis. Similarly, Hcr2-positivc tumors can be treated with drugs that target the Hcr2 axis. For these reasons, the NCCN task force mandate ²² that ER, PR and Hcr2 status be measured for every new case of breast cancer.

[0120] While ER, PR, and Hcr2 status is routinely assessed by immunohistochcmistry (111C) staining for the receptors (Her2 can also be measured by FISH staining), we explored whether H&E morphology, quantified via fingerprints, could serve as a surrogate marker of these proteins. Initially, we queried this hypothesis on a set of breast cancer images curated TCGA. ¹² These samples consist of H&E whole slide images (WSls) from 939 patients at 40 sites. First, scaled the images to 20X resolution, randomly extracted 120 image patches per image, and fed them through the fingerprint network to calculate fingerprints (Figure 7A, step 1 ). Then, a second neural network was trained to compress the fingerprints (512D vectors) into a patch-wise prediction score of a particular receptor from -1 to +1 (Figure 7A, step 2). For simplicity, the figure indicates the process for predicting ER; however, the same procedure was used for PR and Hcr2. Finally, the predictions was averaged across the image to estimate the patient's receptor status (Figure 7A, steps 3-4). We trained on the TCGA dataset with five-fold cross validation. The patients were split into five groups: three were used to train the second network; one was used to monitor training progress and decide when to slop training; the remaining group was tested. The plot in Figure 7B (left) show's the ROC curve for a representative test set from the TCGA data, for ER classification. The average AUC of the test sets w as 0.88 (n=183. per test set). This is the highest HR classification score we have observed, including our previous work using nuclear morphometric features (0.72 AUC) ² ’ and other recent works that predict molecular characteristics of breast cancer from H&E tissue microarray images ^{24 26} . To validate these findings, WSIs of 2531 breast cancers were obtained from the ABCTB ¹³ . and tested whether the TCGA-trained classifier could predict KR-stalus in this group. An AUC of 0.89 on this dataset (n-2531 ) was measured. Applying the same pipeline to PR and Hcr2, it was found that fingerprints could predict PR in the TCGA dataset with an average test set AUC=0.78_(n= 180, per test set), and AUC=0.8I (ABCTB, n-2514) (Figure 7B, center and left). The results for Her2 were AU ^’ C-0.71 (TCGA, n-124) and AUC=0.79 (ABCTB, n=2487). As a methodological control, in addition to these experiments, we trained a classifier to predict ER status directly from image patches and measured an AUC ^~ 0.82 (TCGA, n = 138).

[0121] The fingerprint-based ER classifier learned epithelial patterns

[0122] To understand the regions that were used to make biomarkcr predictions, the trained

ER was applied to whole slides. Our first question was whether the ER classifier worked best on epithelium vs. stroma vs. fat. Briefly, each whole slide image was divided into non-overlapping patches and performed image segmentation to classify each patch as epithelium, stroma, fat. or background. Additionally, a 512D tissue fingerprint for each patch was calculated and predictions made ER for each patch. When the patch-predictions were averaged across the entire slide, they were able to classify ER with an AUC of 0.88 (histogram shown in Figure 8A). The patches were also subset by tissue type and calculated the AUCs after averaging patches of different types (Figure 8B). It w ^r as found that the classifier was most accurate when we pooled predictions from epithelium patches, followed by stroma, followed by fat. Moreover, pooling across epithelium only, or epithelium and stroma was essentially equivalent to pooling across all patches.

[0123] These findings were visually supported by hcatmaps generated from the slides (Figure 80. We present representative slides that were correctly and incorrectly classified: H&E thumbnails, tissue type segmentations, and ER-prcdiction hcatmaps shaded from black to white representing a prediction of -1 to * 1. It was observed that while most areas in the slide have a score of approximately 0 (gray), regions that arc strong predictors (white or black) tend to lie in epithelial regions (green). Even when the predictions arc inaccurate, the classifier makes mistakes based on misclassifying epithelial regions. For instance, false positive and false negative slides, shown on the uppcr-lcfi and low'cr-right quadrants of Figure 8C show strong white and black signal in epithelial regions.

[0124] L low dimensional embedding of fingerprints reveals histologic patterns that predict ER status

[0125] To test the hypothesis that tissues predicted to be KR-positive or negative may have similar visual characteristics, fingerprints from w-hole-slides were embedded into a low dimensional space using a fast implementation of the ISNE algorithm. tSNE is an unsupervised non-linear technique which compresses high dimensional vectors into a low dimensional space, while preserv ing local structures in the data. In our case, we used tSNE to compress a matrix of fingerprints (512D) into a 2D embedding space (Figure 9). Each point in this embedding represents a fingerprint from a single patch, and there are 120 patches (points) per WSI on this plot.

[0126] When the points by the FR score predicted by the ER-classificr were shaded, it was noticed that most points fail to strongly predict ER status (colored green, prediction approx. 0).

However, there w ere several noticeable clusters of points predicted to be ER-negative (blue) or ER- positivc (red). Of these dusters, we manually selected 12 areas to explore (six ER-negative, and six ER-positivc). and present the five patches from each cluster. Of note, each patch is from a different patient. Examination of the image patches reveals shared histologic features within each cluster. For instance, cluster I reveals regions of necrosis and relatively sparse cells. Patches in cluster 2 include patches with large quantities of immune cells, and cluster 7 contains nests of tumor cells.

[0127] DISCUSSION

[0128] Deep learning is transforming computer vision and cancer pathology. As training sets scale, there arc substantial increases in accuracy on diagnosis, prognosis and theragnosis' ⁷ . However, the biggest gains arc likely to come when we leant to leverage a greater spectrum of available pathology images, including the vast majority of images which are mostly or completely unlabclcd. Here, we illustrate a novel first step, using tissue matching to discern features that arc distinctive for a patient but differ between individuals. While tissue matching is not a skill formally taught in pathology training, it allows a neural network to discover key discriminating histologic features from a large set of unannolaicd images. Interestingly, these discriminatory features, or fingerprints, tend to reside at the interfaces between epithelial and stromal cells, and may reflect tissue specific genetic and microenvironmental parameters. While this study used serial sections of a TMA to design a rigorous implementation of core-matching, the resulting network trained in experiment 4 demonstrates that a training paradigm that incorporates style normalization may benefit significantly from histology images of any type, not just matched TMA images.

[0129] Training a network on the task of tissue identification also improves the interprelability ofDNNs and provides insights about the elusive“black box' ^* of deep learning. The ground truth (tissue identify) is indisputable, and visualizations reveal cohesive, biologically intcrprctablc patterns that leverage parameters which arc likely to reflect unique underlying genetic and microenvironmental interactions. Hence, we anticipate that fingerprints, which identify features that discriminate between individuals, will be useful when applied to tasks that seek to discriminate between groups of biologically different tissues.

[0130] The experiments set forth above demonstrate the significant predictive power of such fingerprints to predict the molecular status of a tumor. Taking the fingerprint network, we extracted fingerprint features from whole slide images and used them to predict ER. PR, and Hcr2 status from two independent breast cancer cohorts. We initially trained and validated our algorithms on images from The Cancer Genome Atlas (TCGA), with cross-validation. Then, performed independent validation on samples from the Australian Breast Cancer Tissue bank (ABCTB, n = 2351 ) achieving the following areas under the curve: 0.89 (ER). 0.81 (PR), and 0.79 (Hcr2). These metrics arc higher than all previously published attempts to predict molecular information from H&E images. The improved performance is secondary to the implementation of tissue fingerprinting. The performance we found is similar to previous studies assessing the correlation between II 1C and microarray assessments of ER and PR, which found good concordance between frozen and IHC for ER (93%) and lower for PR (83%).22,28 We believe that using tissue fingerprints will ultimately enable direct treatment response prediction in breast and other cancers, to an accuracy above that provided by current molecular approaches.

[0131] While classification accuracy is an important metric for an algorithm's utility, the significance of fingerprinting extends beyond this, because it enables the interpretation of the histologic patterns learned by deep neural networks. At present, if inlcrprctabilily is the goal, deep learning is not necessarily the best approach. Using human designed, handcrafted, pro-extracted features such as cell shape, cell neighborhood statistics can provide rapidly intcrprctablc insights about the tissue properties that correlate to a clinical or molecular outcome .29.30 However, the downside of these approaches is introduction of human bias and the challenge of building workflows to accurately extract these features.

[0132] While the flexibility and aulomaiicity of deep learning makes it effective for black box usage in a number of scenarios, intcrprciability is essential for developing testable hypotheses that advance basic biomedical research. Thus, we were encouraged by the intcrprciability of our fingerprint -based ER classifier. The prediction hcatmaps shown in Figure 8C demonstrate that the network learned to make positive, negative, and neutral predictions. Thus, it automatically learned to ignore some regions (c.g. neutral, gray areas), while paying attention to others. In this case, it learned to pay attention to areas of tissue epithelium.

[0133] A second insight came from plotting the ISNF. embedding of fingerprints and discovering natural clusters of patches with similar histologic characteristics predicted to be ER- positivc, ER-ncgative, or neutral (Figure 9). There is a significant history of histological classification of breast cancer patterns. Numerous attempts have been made to develop histologic typing schemes, but these arc subjective, difficult to reproduce, and a large number of slides frequently fall into the“no specified type" category (lacking distinctive characteristics of any pre-dcllncd type). The embedding we present in Figure 9 provides preliminary support for the development of machine-automated histologic typing, in which tissue patches are clustered by visual histologic similarity. Of note, our algorithm learned this embedding without knowing the molecular status of the tissues, and the embedding is not sensitive to differences in staining colors. Patches in these clusters have different shades of pink and purple H&E but share core histologic features.

[0134] The observation that some of the clusters seem to correlate w'ith ER status suggests a link between tissue architecture and clinical ER status. This view suggests that only a subset of patches tissue architectures contain useful diagnostic information for making this prediction. Future work will determine whether predicted fingerprints can be used to make biomarkcr predictions while providing confidence scores based on the quantities and types of tissues present in the slide. Assigning a confidence score to a prediction may facilitate triage and workup of the most relevant biological markers for a particular patient.

| [0135] An additional area to focus on is how these visualizations can be used to improve a classifier with intemiediale accuracy. A potential limitation of this approach is that we demonstrate its application with respect to ER classification, but not PR or HER2. We made this decision believing that focusing on the ER classifier, w hich had the highest accuracy, would reduce the chances of misinterpreting the visualizations. However, based on our findings, we believe that using these visualization techniques may provide diagnostic information to troubleshoot difficult classification questions.

[0136] While exemplary embodiments arc described above, it is not intended that these embodiments describe all possible forms of the invention. Rather, the w ords used in the specification arc words of description rather than limitation, and it is understood that various changes may be made without departing from the spirit and scope of the invention. Additionally, the features of various implementing embodiments may be combined to form further embodiments of the invention. [0137] REFERENCES

[0138] 1. Krizhcvsky, A., Sutskcvcr, 1. & Hinton. G. E. ImagcNct Classification with

Deep Convolutional Neural Networks.

[0139] 2. Ehteshami Bejnordi. B. ct al. Diagnostic Assessment of Deep learning

Algorithms for Detection of Lymph Node Mclaslases in Women With Breast Cancer. JAMA 318, 2199 (2017).

]0140] 3. Esicva, A. et al. Dermatologist-level classification of skin cancer with deep neural networks. Nature 542, 1 15-1 18 (2017).

[0141] 4. Tcarc. P.. Fishman. M„ Benzaquen, O.. Toledano, E. & Elnckavc. E.

Malignancy Detection on Mammography Using Dual Deep Convolutional Neural Networks and Genetically Discovered False Color Input Enhancement. J. Digit. Imaging 30, 499-505 (2017).

[0142] 5. Liu, Y. ct al. Artificial Intelligence-Based Breast Cancer Nodal Metastasis

Detection. Arch. Pathol. Lab. Med. (2018). doi:10.5858-arpa.20l8-0147-oa

[0143] 6. Beck, A. H„ Irshad. H., Gargeya, R„ Khosla, A. & Wang, D. Deep learning

Based Cancer Mctastascs Detection.

[0144] 7. Coudray, N. ct al. Classification and mutation prediction from non-small cell lung cancer hislopathology images using deep learning. Nat. Med. 24. 1559 1567 (2018).

[0145] 8. Bychkov, D. et al. Deep learning based tissue analysis predicts outcome in colorectal cancer. Sci. Rep. 8, 3395 (2018).

[0146] 9. Allison. K. H. et al. Understanding diagnostic variability in breast pathology:

Lessons learned from an expert consensus review panel. Hislopathology 65, 240-251 (2014).

[0147] 10. Elmore, J. G. ct al. Diagnostic Concordance Among Pathologists Intciprcting

Breast Biopsy Specimens. 98104, 1 122-1132 (2017). [0148] I I . Robbins, P. et al. Histological grading of breast carcinomas: A study of intcrobscrvcr agreement. Hum. Pathol. 26, 873-879 (1995).

[0149] 12. Koboldt, D. C\ ct al. Comprehensive molecular portraits of human breast tumours. Nature 490, 61 -70 (2012).

[0150] 13. Carpenter, J., Marsh, D„ Mariasegaram, M. & Clarke. C. The Australian Breast

Cancer Tissue Bank (ABCTB). Open J. Biorcsour. 1, cl (2014).

]0151] 14. Hammond, M. E. H. ct al. American society of clinical oncology college of

American pathologists guideline recommendations for immunohisiochcmical testing of estrogen and progesterone receptors in breast cancer (unabridged version). Arch. Pathol. Lab. Med. 134. (2010).

[0152] 15. He, K-, Zhang, X., Ren, S. & Sun, J. Deep Residual Learning for Image

Recognition.

[0153] 16. Paszkc, A. ct al. Automatic differentiation in PyTorch. 31st Conf. Neural Inf.

Process. Syst. 1 -4 (2017). doi: 10.1017 CB09781 107707221.009

[0154] 17. Zhu, J.-Y., Park, T„ I sola, P. & Efros, A. A. Unpaired Imagc-to-lmagc

Translation using Cycle-Consistent Adversarial Networks. Proc. IEEE Int. Conf. Comput. Vis. 2017. 2242-2251 (2017).

[0155] 18. Kathcr. J. N. ct al. Predicting survival from colorectal cancer histology slides using deep learning: A retrospective multiccntcr study. PLOS Med. 16, el 002730 (2019).

[0156] 19. Lindcrman. G. C„ Rachh. M., Hoskins, J. G., Stcincrbcrgcr, S. & Klugcr, Y.

Fast interpolation-based t-SNE for improved visualization of single-cell RNA-scq data. Nat. Methods

16, 243-245 (2019).

[0157] 20. Gatys. L. A., Eckcr, A. S. & Bcthgc. M. A Neural Algorithm of Artistic Style.

(20! 5k [0158] 21. Van Der Maalen, I„ & Hinton, G. Visualizing Data using l-SNF. Journal of

Machine Learning Research 9, (2008).

)01S9) 22. Allred, D. ct al. NCCN Task Force Report: Estrogen Receptor and Progesterone

Receptor Testing in Breast Cancer by Immunohistoehcmistry. J Natl Conipr Cane Netw 7, 1 -21 (2009).

[0160] 23. Rawat. R. R., Rudcrman, D.. Macklin, P., Rimm, D. L. & Agus. D. B.

Correlating nuclear morphometric patterns with estrogen receptor status in breast cancer pathologic specimens, npj Breast Cancer 4, 32 (2018).

)0161) 24. Couture, H. D. ct al. Image analysis with deep learning to predict breast cancer grade, FR status, histologic subtype, and intrinsic subtype, npj Breast Cancer 4. 30 (2018).

[0162] 25. Shamai, G. ct al. Artificial Intelligence Algorithms to Assess Hormonal Status

From Tissue Microarrays in Patients With Breast Cancer. 2, I— 1-4 (2019).

[0163] 26. Jabcr. M. I. ct al. A deep learning image-based intrinsic molecular subtype classifier of breast tumors reveals tumor heterogeneity that may affect survival. I 10 (2020).

[0164] 27. Campanella. G., Silva, V. W. K. & Fuchs, T. J. Terabyte-scale Deep Multiple

Instance Learning for Classification and Localization in Pathology. (2018).

[0165] 28. Rocpman, P. ct al. Microarray-Based Determination of Estrogen Receptor,

Progesterone Receptor, and HKR2 Receptor Status in Breast Cancer. Clin. Cancer Res. 15, 7003 701 1

(2009).

)0166| 29. Yu, K.-H. ct al. Predicting non-small cell lung cancer prognosis by fully automated microscopic pathology image features. Nat. Comniun. 7, 12474 (2016).

[0167] 30. Corrcdor, G. ct al. Spatial architecture and arrangement of tumor-infiltrating lymphocytes for predicting likelihood of recurrence in early-stage non-small cell lung cancer. Clin,

Cancer Res. clincanres.20l 3.2018 (2018). doi: I0.1 1 S8 I078-0432.CCR-I8-20I3.