System and method for determining tumor cellularity

Abstract

A system and method for determining tumor cellularity in a tissue sample are provided. In some aspects, the method includes obtaining imaging data associated with a tissue sample. The method also includes predicting, based on an application of a first trained artificial intelligence (AI) model configured to predict at least one target tissue in a region of interest (ROI) of the tissue sample to the imaging data, a predicted target tissue within the ROI of the tissue sample. The method also includes predicting a predicted cellular component count of at least one cellular component within the predicted target tissue. The method also includes determining, based on the predicted cellular component count, data indicative of a tumor cellularity in the tissue sample.

Description

SYSTEM AND METHOD FOR DETERMINING TUMOR CELLULARITYCROSS-REFERENCE TO RELATED APPLICATIONS

[0001] This application claims the benefit of, and priority to, U.S. Provisional Patent Application No. 63 / 746,868 filed on January 17, 2025, which is hereby incorporated by reference herein in its entirety.TECHNICAL FIELD

[0002] The present disclosure relates generally to detecting and treating disease, and more specifically to a system and method for determining tumor cellularity.BACKGROUND

[0003] Tumor cellularity (TC), which may be defined as the ratio of abnormal or neoplastic cells to the total number of cells in a tissue sample, is a key parameter that may aid in identifying clinically relevant mutations that can inform diagnosis, prognosis, and / or clinical management of patients with certain medical conditions. Tumor cellularity is particularly important for assessing the condition of patients with various cancers, such as lung adenocarcinoma (LUAD). For instance, molecular pathology reports routinely involve TC estimation to help identify relative sensitivity in testing samples for tumor-specific molecular alterations. While TC assessment is routinely performed prior to molecular testing, there are currently no best practice or consensus guidelines for determining TC.

[0004] In current practice, tumor cellularity is often estimated visually by a pathologist using hematoxylin and eosin (H&E) stained slides, sometimes supplemented by manual cell counting in selected regions of interest. These manual approaches are imprecise, labor-intensive, and subject to substantial inter- and intra-ob server variability, particularly in specimens with certain complex tumor architectures. Some digital approaches have been employed, but these techniques often are imprecise and require high levels of human oversight and annotation to correct errors. This annotation burden is time-consuming, limits scalability, and may still be affected by subjectivity and variability in how different annotators delineate tumor and non-tumor sections of the image. Current computer-implemented methods may also classify benign cells, stromal regions, macrophages, or pigmented artifacts as tumor, or may fail to identify certain tumor structures, thereby biasing tumor cellularity estimates.14913-5900-09642065472-001009 WOPT

[0005] Thus, there is a need in the art for improved systems and methods for determining TC in tissue samples in an accurate, efficient, and automated manner using artificial intelligence processes.SUMMARY

[0006] According to some implementations of the present disclosure, a method for determining tumor cellularity in a tissue sample is provided. In some aspects, the method includes obtaining, by at least one processor, and via a communications interface, imaging data associated with a tissue sample. The method also includes predicting, by the at least one processor, and based on an application of a first trained artificial intelligence (Al) model configured to predict at least one target tissue in a region of interest (ROI) of the tissue sample to the imaging data, a predicted target tissue within the ROI of the tissue sample. The method also includes predicting, by the at least one processor, and based on an application of a second trained Al model to the imaging data, a predicted cellular component count of at least one cellular component within the predicted target tissue. The method also includes determining, by the at least one processor, and based on the predicted cellular component count, data indicative of a tumor cellularity in the tissue sample.

[0007] According to some implementations of the present disclosure, a system for determining tumor cellularity in a tissue sample is provided. In some aspects, the system comprises a communications interface configured to obtain imaging data associated with a tissue sample, a storage device configured to store the imaging data, a memory storing instructions, and at least one processor communicatively coupled with the memory and configured to execute the instructions. The at least one processor is configured to predict, based on an application of a first trained artificial intelligence (Al) model configured to predict at least one target tissue in a region of interest (ROI) of the tissue sample to the imaging data, a predicted target tissue within the ROI of the tissue sample. The system is further configured to predict, based on an application of a second trained Al model to the imaging data, a predicted cellular component count of at least one cellular component within the predicted target tissue. The system is further configured to determine, based on the predicted cellular component count, data indicative of a tumor cellularity in the tissue sample.

[0008] The above summary is not intended to represent each implementation or every aspect of the present disclosure. Additional features and benefits of the present disclosure are apparent from the detailed description and figures set forth below.24913-5900-09642065472-001009 WOPTBRIEF DESCRIPTION OF THE DRAWINGS

[0009] The disclosure, and its advantages and drawings, may be understood from the following description of representative embodiments together with reference to the accompanying drawings. The drawings depict only representative embodiments, and should not be considered as limitations on the scope of the various embodiments or claims.

[0010] FIG. l is a block diagram of an example system, according to aspects of the present disclosure;

[0011] FIG. 2 is a flowchart setting forth steps of a method, according to aspects of the present disclosure;

[0012] FIG. 3 is a graphical illustration showing steps of a developing a pipeline, according to aspects of the present disclosure.

[0013] FIG. 4A is a graphical illustration showing steps of a method, according to aspects of the present disclosure.

[0014] FIG. 4B is another graphical illustration showing steps of a method, according to aspects of the present disclosure.

[0015] FIG. 4C is another graphical illustration showing steps of a method, according to aspects of the present disclosure.

[0016] FIG. 5 A is a graphical illustration comparing manual identification with an artificial intelligence (Al) approach, according to aspects of the present disclosure.

[0017] FIG. 5B is another graphical illustration comparing manual identification with an artificial intelligence (Al) approach, according to aspects of the present disclosure.

[0018] FIG. 5C is yet another graphical illustration comparing manual identification with an artificial intelligence (Al) approach, according to aspects of the present disclosure.

[0019] FIG. 6A is a graphical illustration showing example regions of interest (ROIs) in images of stained tissue and prediction maps, according to aspects of the present disclosure.

[0020] FIG. 6B is another graphical illustration showing example regions of interest (ROIs) in images of stained tissue and prediction maps, according to aspects of the present disclosure.

[0021] FIG. 6C is another graphical illustration showing example regions of interest (ROIs) in images of stained tissue and prediction maps, according to aspects of the present disclosure.

[0022] FIG. 6D is another graphical illustration showing example regions of interest (ROIs) in images of stained tissue and prediction maps, according to aspects of the present disclosure.

[0023] FIG. 7A is a graphical illustration showing example masks outputted, according to aspects of the present disclosure.34913-5900-09642065472-001009 WOPT

[0024] FIG. 7B is another graphical illustration showing example masks outputted, according to aspects of the present disclosure.

[0025] FIG. 7C is another graphical illustration showing example masks outputted, according to aspects of the present disclosure.

[0026] FIG. 8A is a graph showing tumor cellularity (TC) assessments, and effect of TC thresholding on the number of on samples selected for molecular testing by raters, according to aspects of the present disclosure.

[0027] FIG. 8B is another graph showing tumor cellularity (TC) assessments, and effect of TC thresholding on the number of on samples selected for molecular testing by raters, according to aspects of the present disclosure.

[0028] FIG. 8C is another graph showing tumor cellularity (TC) assessments, and effect of TC thresholding on the number of on samples selected for molecular testing by raters, according to aspects of the present disclosure.

[0029] FIG. 8D is another graph showing tumor cellularity (TC) assessments, and effect of TC thresholding on the number of on samples selected for molecular testing by raters, according to aspects of the present disclosure.

[0030] FIG. 9A is a graphical illustration showing example ROIs with stained tissue, according to aspects of the present disclosure.

[0031] FIG. 9B is another graphical illustration showing example ROIs with stained tissue, according to aspects of the present disclosure.

[0032] FIG. 9C is another graphical illustration showing example ROIs with stained tissue, according to aspects of the present disclosure.

[0033] FIG. 10A is a graphical illustration showing an example immunohistochemistry (IHC) stained slide and ROI, according to aspects of the present disclosure.

[0034] FIG. 10B is a graphical illustration showing a detailed view of an example immunohistochemistry (IHC) stained slide and ROI, according to aspects of the present disclosure.

[0035] FIG. 10C is a graphical illustration showing another detailed view of an example immunohistochemistry (IHC) stained slide and ROI, according to aspects of the present disclosure.

[0036] FIG. 11A is a graphical illustration showing accuracy metrics as confusion matrices, and demonstrating performance of an artificial intelligence model, according to aspects of the present disclosure.

[0037] FIG. 1 IB is a graph showing a boxplot demonstrating performance of an artificial 44913-5900-09642065472-001009 WOPTintelligence model, according to aspects of the present disclosure.

[0038] FIG. 12A is a graphical illustration showing example hematoxylin and eosin (H&E) and IHC ROIs and masks predicted by an artificial intelligence approach, according to aspects of the present disclosure.

[0039] FIG. 12B is another graphical illustration showing example hematoxylin and eosin (H&E) and IHC ROIs and masks predicted by an artificial intelligence approach, according to aspects of the present disclosure.

[0040] FIG. 12C is another graphical illustration showing example hematoxylin and eosin (H&E) and IHC ROIs and masks predicted by an artificial intelligence approach, according to aspects of the present disclosure.

[0041] FIG. 12D is another graphical illustration showing example hematoxylin and eosin (H&E) and IHC ROIs and masks predicted by an artificial intelligence approach, according to aspects of the present disclosure.

[0042] FIG. 12E is another graphical illustration showing example hematoxylin and eosin (H&E) and IHC ROIs and masks predicted by an artificial intelligence approach, according to aspects of the present disclosure.

[0043] FIG. 12F is another graphical illustration showing example hematoxylin and eosin (H&E) and IHC ROIs and masks predicted by an artificial intelligence approach, according to aspects of the present disclosure.

[0044] FIG. 12G is another graphical illustration showing example hematoxylin and eosin (H&E) and IHC ROIs and masks predicted by an artificial intelligence approach, according to aspects of the present disclosure.

[0045] While specific implementations and embodiments are shown in the drawings and described in detail herein by way of example, various modifications and alternative forms are possible. Hence, it should be understood that the present disclosure is not intended to be limited to particular forms disclosed. Rather, the present disclosure covers all modifications, equivalents, and alternatives falling within the spirit and scope of the present disclosure as defined by the appended claims.DETAILED DESCRIPTION

[0046] Tumor cellularity (TC), defined as a ratio of abnormal or neoplastic cells to a total number of cells in a sample, is typically assessed as an average “eyeball” estimate from multiple regions of a slide and / or by manual cell count. Such approach is imprecise, impractical, labor-intensive, and subject to confounding factors, such as tissue shrinkage,54913-5900-09642065472-001009WOPTdeformation and fixation quality, tumor cell size, density and distribution, tumor growth pattern and pattern heterogeneity, as well as extent of tumor invasion and admixture of tumor cells with non-tumor cells (epithelial, immune, stromal, endothelial, and others) in the tumor microenvironment. Hence, substantial interobserver variability has been reported with gestalt and manual TC estimates. Also, studies using hematoxylin and eosin (H&E) stained slides have shown that “eyeball” gestalt assessment overestimates TC when compared with manual cell counting. Further, overestimation of TC in samples submitted for molecular testing can be an important root cause of false negative molecular test results. Failure to detect the presence of clinically actionable molecular alterations can result in delayed treatment, failure to receive optimal treatment, and / or inappropriate exclusion from one or more clinical trials while also wasting financial resources of patients and laboratory.

[0047] It is recognized herein that digitization of tissue slides when combined with artificial intelligence (Al) offers potential solutions to conventional practice. In one approach, after selecting regions of interest (ROIs) from digital H&E stained slides of LUAD, a commercial digital pathology platform with deep learning (DL) was used to run two DL models (one to automatically delineate areas with cancer cells and the other to detect cell nuclei in the ROIs). To assess TC, detected nuclei were counted within the ROIs and within the cancer area identified in the ROIs. Prior to analyzing the ROIs, a pathologist trained a model for cancer cell delineation using manually annotated image tiles. Insufficient representation of tumor in the training set slides resulted in the trained model having limited accuracy in detecting cancer cells; therefore, the assessed TC was adjusted to account for false positive and false negative cancer cell detections. In another approach, classification and regression models were applied to assess TC in ROIs. The classification model categorized ROIs as either devoid of or containing cancer cells. The ROIs identified as containing cancer cells were then further analyzed by a regression model that predicted TC in those ROIs. Training data for model developments were prepared by a pathologist who manually counted cancer and non-cancer cells in the ROIs with cancer. In another approach, TC assessment involved manual circling of areas with tumor cells in whole slide images (WSIs) using an open-source software followed by recognition and counting cancer and non-cancer cells within the circled areas by a cell classifier built into the software. The classifier was trained using cells manually annotated by the pathologist. Since cell classification accuracy varied between the analyzed WSIs, the classifier required retraining for some WSIs to obtain more reliable estimates of the cancer cell counts.64913-5900-09642065472-001009WOPT

[0048] Substantial involvement of pathologists in developing and / or running TC assessment algorithms is a common disadvantage of these approaches. For model training, making single-cell level annotations, labeling cells, and circling tumors in individual H&E slides or ROIs by hand is subjective, labor-intensive, and has low throughput. Cell confluency and tumor microenvironment composition that vary across slides and additionally affect the accuracy and reproducibility of manual annotations. Although many LUAD specimens submitted for molecular testing include small samples (biopsies and / or cytology cell blocks), a substantial proportion of molecular testing is performed on surgical resection specimens (primary tumor and / or metastatic lesions).

[0049] In particular, the present disclosure improves upon conventional approaches by providing a system and method for determining TC using imaging. In some aspects, an AI-based processing pipeline for the assessment of TC using imaging data is utilized, providing a number of advantages, benefits, and improvements to detecting and treating disease. For instance, the present approach addresses the pressing need for more objective, accurate, reproducible, and cost- effective approach to determining TC in tissue samples (e.g., surgical resection specimens).

[0050] FIG. 1 shows a block diagram of an example system 100, according to aspects of the present disclosure. The system 100 may include one or more processing devices 110 configured to carry out steps in accordance with aspects of the present disclosure. In general, a processing device 110 may include one or more processors 112, one or more memories 114, one or more output devices 116, one or more input devices 118, or a combination thereof, as well as other components. For example, a processing device 110 may include a computer, a laptop, a tablet, a smart device (e.g., smart phone), a mobile device (e.g., mobile telephone, personal digital assistant (PDA), etc.), a server, a mainframe, a cloud computing system, a wearable device, and so forth. In some embodiments, the processing device(s) 110 may be incorporated into a monitoring and / or treatment device, system, or apparatus. For example, the processing device 110 may be integrated into a control system or an interface system of a treatment system, device, or apparatus.

[0051] A processor 112 can include any processing device, such as a computer processing unit (CPU), graphical processing unit (GPU), microprocessor, digital signal processor, microcontroller, application specific integrated circuit (ASIC), programmable logic device (PLD), field programmable logic devices (FPLD), programmable gate array (PGA), field programmable gate array (FPGA), and so forth. The processor(s) 112 may be configured to carry out various tasks to operate a processing device 110.74913-5900-09642065472-001009 WOPT

[0052] The processor(s) 112 may be configured to implement a processing pipeline to carry out various steps, according to this disclosure. In some embodiments, the processor(s) 112 may be configured to generate, train, and / or utilize various Al models, in accordance with aspects of the present disclosure. For instance, the processor(s) 112 may be configured to generate, train, and / or utilize a first Al model configured to predict various target tissues in a region of interest (RO I) of the tissue sample, a second Al model configured to predict a count of various cellular components in targeted tissue, or both. In some embodiments, the processing pipeline may also include steps to determine TC, as described further herein. In further embodiments, the processor(s) 112 may be configured to generate, train, utilize, and / or apply a third Al model configured to determine, generate, modify, or otherwise establish a treatment recommendation based on the determined TC, the target tissues, or any other input data. The third trained Al model may be applied to data indicative of a tumor cellularity of the tissue sample. The third trained Al model may also be applied to other data associated with the tissue sample. The third trained Al model may be a trained machine learning model. For example, the first, second, or third trained Al model may be a (i) a decision tree, (ii) a Bayesian network, (iii) an artificial neural network, (iv) a support vector machine, (v) a convolutional neural network, (vi) a capsule network, or (vii) any combination of (i)-(vi). The established treatment recommendation represents a treatment for the individual determined based on the determined TC, the target tissues, or any other input data. The treatment recommendation may be intended to treat, manage, or otherwise affect a condition of the individual or the tumor of the target tissue.

[0053] The memory 114 can include any suitable memory device and / or machine-readable medium capable of storing, encoding, and / or carrying a set of instructions for execution by a processing device and that cause the processing device to perform and / or implement any of the features discussed herein, including solid-state memories, optical media, magnetic media, random access memory (RAM), read only memory (ROM), a floppy disk, a hard disk, a CD ROM, a DVD ROM, flash memory, or other computer readable medium that is read from and / or written to by a magnetic, optical, or other reading and / or writing system that is coupled to the processing device, can be used for the memory or memories. In some embodiments, the memory 114 may include machine-readable instructions for carrying out various steps of methods disclosed herein, and / or other methods. The memory 114 can also store data associated with steps of methods described herein.

[0054] The output device 116 may include various output devices, such as any type of display device (e.g., monitor, touchscreen, and so forth). A display device can include any 84913-5900-09642065472-001009WOPTdisplay technology, including but not limited to one or more display devices using Liquid Crystal Display (LCD) or Light Emitting Diode (LED) technology. In some implementations, an output device 116 may be used to display or report any data and information associated with the features disclosed herein, including the results of analysis and / or prediction data or output generated by various Al models, recommendations, and so forth, according to description herein. For instance, in some embodiments, the output device 116 may be configured to provide a report indicative of a determined TC, according to the present disclosure. The output device 116 may also provide a report indicating a recommendation (e.g., for treatment, further steps, such as molecular testing, and so forth), according to determined TC.

[0055] In some implementations, an output device 116 may be used to communicate, provide, and / or store various data, information, and / or signals, for instance, via a wired and / or wireless communication network. For example, in some implementations, the processor(s) and / or output device 116 may generate one or more signals, based on prediction data or output generated by various Al models, as well as data and information derived therefrom, and transmit the signal(s) to a device, system, apparatus, and so forth, configured to carry out various tasks responsive to such received signal(s), such as initiating, pausing, terminating, and / or recommending a treatment, and so forth.

[0056] In some implementations, a communications interface 119 may be used to communicate, provide, and or store various data, information, and / or signals, for instance, via a wired and / or wireless communication network. The communications interface 119 may be used to communicatively couple the system 100, the processing device 110, and / or the data repository 120 with each other, as well as with a wired or wireless communications network. The communications interface 119 may also be used to communicatively couple the processing device 110 with external computing systems such as a computing server system or a cloud computing network.

[0057] The input device 118 may include various input devices, such as any type of mouse, keyboard, microphone, camera, and so forth. In some embodiments, the input device 118 may include at least one electronic interface configured to receive various data associated with the individual. The input device 118 may also include a user interface that may allow a user to interact with the system 100 and / or processing device 110 for any suitable purpose, including providing user input, initiating, pausing, and / or terminating analysis, training, and / or execution of an Al model, adjusting one or more parameters of analysis, model, and so forth.

[0058] In some embodiments, the system 100 may include data repository 120. The data repository 120 may be operated using various systems, devices, computers, servers, databases,94913-5900-09642065472-001009WOPTand other hardware. In some embodiments, the data repository 120 may store imaging data 122, along with various other data and information, such as patient data.

[0059] As illustrated in FIG. 1, the data repository 120 may be connected or connectable to a processing device 110. As such, the processing device 110 (and / or other processing devices or processors) may store data in and / or access data from the data repository 122. For instance, in some implementations, the processing device 110 may utilize imaging data to train an Al model, according to aspects of the present disclosure. In some implementations, the processing device 110 may store prediction data, for instance, indicative of target tissue(s) as well as count of various cellular components, as described herein.

[0060] In some embodiments, the processing device 110 may be connected or connectable, via wired and / or wireless communication, to a treatment apparatus 124, such as an apparatus configured to administer a treatment to a patient. For example, the treatment apparatus 124 could be, but is not limited to an infusion pump, a chemotherapy infusion system, a radiation therapy delivery system, a radiofrequency system, a focused ultrasound therapy system, a surgical system, an immunotherapy infusion system, a targeted molecular therapy delivery system, or another type of treatment apparatus. Responsive to, for instance predicted data, as well as data and information determined therefrom (e.g., TC), the processing device 110 and / or output device 116 may be configured to generate and transmit one or more signal to the treatment apparatus 124. Responsive to the transmitted signal(s), the treatment apparatus 124 may initiate, modify, pause, and / or cease a treatment to a patient.

[0061] Turning now to FIG. 2, a flowchart setting forth steps of a process 200, according to aspects of the present disclosure is illustrated. The process 200 may be carried out using any suitable device, apparatus, or system, such as the system 100 described with reference to FIG. 1. In some embodiments, various steps of the process 200 may be implemented as instructions stored in non-transitory computer-readable media, as a program, firmware or software, and executed by various general-purpose, programmed or programmable computers, processors or other processing devices. In other embodiments, various steps of the process 200 may be hardwired in an application-specific computer, server, processor, dedicated system, or module.

[0062] The process 200 may begin at process block 202 with obtaining imaging data, as well as other data and information, associated with associated with a tissue sample, for instance, obtained from an individual or patient. By way of example, the imaging data may include one or more images, or portions thereof, depicting one or more tissue samples or selected regions104913-5900-09642065472-001009 WOPTwithin the sample stained using various markers, as described further herein. The tissue sample may be sourced from various biological sources. For example, the tissue sample may be a biopsy slide, or a portion of a biopsy slide. The biopsy slide may include a biopsy specimen collected from a tissue of the individual or patient. For example, the biopsy specimen may be a biopsy specimen collected from lung tissue of the individual. Other types of tissue, such as surgical resection specimens, cell specimens, fluid specimens, and other specimens, may be included in the tissue sample. In some implementations, imaging data may be accessed at process block 202 from a data repository, device, system, database, memory, or other data storage location. Received and / or accessed imaging data may undergo various processing steps, as described further herein. For example, imaging data, or portions thereof, may undergo image segmentation, image filtering, image registration, intensity thresholding, and so forth.

[0063] As indicated by process block 204, received, accessed, and / or processed imaging data may then be provided as input to a first Al model, where the first Al model may be configured (e.g., using various training techniques) to predict one or more target tissue in a ROI of a tissue sample or in the tissue sample (e.g., the entire tissue sample). For example, the first Al model may be configured to predict tumor tissue, normal tissue, benign tissue, stroma tissue, and so forth. In some implementations, one or more tissue mask or tissue maps may be generated at process block 204 using image pixels corresponding to the target tissue(s) predicted using the first Al model. For example, first Al model may include a convolutional neural network or other deep learning architecture trained to perform semantic segmentation and to assign, on a pixel-wise or patch-wise basis, labels indicative of tumor tissue, benign or normal epithelium, stroma, necrotic regions, and / or other tissue categories of interest. The first Al model may also generate one or more tissue masks or tissue maps that classify pixels or voxels into malignant epithelium, benign epithelium, and stromal tissue classes, as well as other classes of tissues. In other example embodiments, the tissue masks define a spatial extent, shape, and / or relative location of predicted target tissues within the ROI or tissue sample for use in subsequent cellular component counting and tumor cellularity determination.

[0064] Imaging data may then be provided at process block 206 to a second Al model configured (e.g., using various training techniques) to predict count of various cellular components (e.g., a cell nucleus, and so forth) in target tissue(s) predicted using the first Al model, as indicated by process block 206. In some implementations, the second Al model may be applied to the tissue mask(s) or tissue map(s) generated at process block 204. In some implementations, the second Al model operates as a nuclei detection and segmentation step that outputs data associated with cell nuclei across the ROI or tissue sample based on the 114913-5900-09642065472-001009WOPTdetermined target tissues of process block 204. In some implementations, the second Al model may also be applied to imaging data associated with the ROI or tissue sample to predict a total count of cells in the ROI or tissue sample. The second Al model may be applied selectively within pixels classified as malignant epithelium, benign epithelium, and / or stroma in the tissue mask, enabling computation of both a malignant-cell count (for example, nuclei within malignant epithelium regions) and a total cell count (for example, all detected nuclei within the ROI or whole slide image).

[0065] As indicated by process block 208, a TC may then be determined in the tissue sample, according to the present disclosure. For instance, in some implementations, TC may be determined by using the count of the cellular component(s) predicted and the total count of cells in the ROI or tissue sample. For instance, a ratio of a number of tumor cells (determined based on a count of cell nuclei associated with predicted tumor tissue), and a total count of cells in the ROI or tissue sample may be computed to determine tumor cellularity in the ROI or tissue sample.

[0066] In some implementations, a report may be generated, as indicated by process block 210. Process block 210 may also not be performed, in some example embodiments of the present disclosure. The report may be in any form, and include any data and information. In one example, a report may include an indication of a determined tumor cellularity in a tissue sample or ROI associated with a tissue sample. In another example, a report may include an indication of malignant or tumor tissue (e.g., as an image, overlay, map, mask, and other representation) in an ROI or tissue sample, for instance. The report may be provided intermittently, periodically, and / or subject to user input or request. In some implementations, the report may include one or more recommendations. The recommendation may be a recommendation for a treatment or therapy administered by a treatment apparatus. For example, process block 210 may include generating a treatment recommendation configured to initiate, modify, pause, or cease a treatment performed by treatment apparatus 124. In other implementations, the report may include one or more signals executable or actionable by, for instance, a treatment apparatus, imaging device or system, and so forth. To this end, various recommendation(s), signal(s) may be generated and transmitted to various receivers, such as a clinician, treatment apparatus, and so forth. The signal may include data that causes a treatment apparatus or system to initiate, modify, pause, or cease a treatment performed by the treatment apparatus or system.

[0067] Although the process 200 is illustrated and described as a sequence of steps, it is contemplated that the steps may be performed in any order or combination, need not include 124913-5900-09642065472-001009WOPTall illustrated steps, and may include additional steps. The process 200 and / or steps therein may be carried out any number of times, such as intermittently, periodically, and / or subject to user input.

[0068] By way of example, a study illustrating aspects of the present disclosure was performed. In the study, an Al-based pipeline was developed to assess TC in whole slide hematoxylin and eosin stained (H&E) images (WSIs) and in tumor areas (TAs) within WSIs. The pipeline included using a CaBeSt-Net model trained to mask cancer cells, benign epithelial cells, and stroma in H&E WSIs using slides restained with IHC stains, and a model to detect all cell nuclei. High masking accuracy (>91%) by CaBeSt-Net computed using 1024 H&E ROIs and intraclass correlation coefficient ICOO.97 assessing reliability of TC assessments by one pathologist and Al in 20 test ROIs supported pipeline’s applicability to TC assessment in 50 study H&E WSIs. Utilizing the pipeline, TCs assessed in TAs and WSIs were compared with those by 3 pathologists. Reliabilities of TC ratings by the pathologists and the Al pipeline were moderate (ICC>0.48, p<0.0001). Applying TC<20% cut-point to categorize each WSI as inadequate or adequate for molecular testing yielded slight agreement (p=0.1) among the 3 pathologists. Including Al as an additional rater improved sample categorization (moderate, p<0.0001). TC agreements in TAs were consistently higher than in corresponding WSIs, and the consistency and statistical significance held despite various cut-points. As appreciated from description herein, the present pipeline can aid pathologists to assess TC objectively and more reproducibly.

[0069] For method development, H&E sections restained using immunohistochemistry (IHC) with antibodies reactive to proteins in epithelial cells were used to obtain surrogate yet fine annotations of malignant cells. Pipeline components and validation in several datasets demonstrating the pipeline’s tumor cell detection accuracy is described further herein. In one application, the present pipeline was applied to assess TC in 50 H&E WSIs and in an arbitrarily selected TA within each WSI. Statistical analyses of TCs assessed and by the pipeline summarize the pipeline's applicability in qualifying samples for molecular testing.

[0070] Development of a method for the study to compare manual and Al-based TC assessments in WSIs with LUAD is illustrated in graph 300 of FIG. 3. In particular, H&E and IHC stained slides were to train two DL models. Block 302 shows how a first DL model (Ep-Net) was developed to delineate epithelium in ROIs from the IHC stained slides. Details on the model shown in block 302 are also included in Table 1. The Ep-Net model of block 302 was trained using IHC ROIs (red marker) to output epithelial masks (green). Epithelial masks outputted by Ep-Net were then used to train a second DL model at block 304 (CaBeSt-Net).134913-5900-09642065472-001009WOPTThe second DL model of block 304 may be configured to recognize and provide masks of malignant epithelium (CA), normal / benign epithelium (BE) and stroma (ST) in H&E ROIs and WSIs from a study set (right panel in FIG. 3). In some example embodiments, the first and the second DL models are based on the architecture of DeepLabV3+ convolutional neural network pre-trained on ImageNet database. Any other model and pretraining method may also be used to establish the first and the second DL model, including models that utilize Al techniques other than deep learning. After testing, the CaBeSt-Net model was incorporated into an Al pipeline for TC assessment. The TC assessments provided by the Al pipeline and were compared. The pipeline includes StarDist - a DL model to detect all nuclei (yellow rims in FIG.3). As shown, tumor area (TA) may be highlighted within a WSI.Table 1. Digital slide and ROI characteristics. The CaBeSt-Net model was tested on ROIs with manual (a,c) and H4C- based (b) ground truth masks. ROIs marked with (*) were extracted from the study set H&E WSIs.&& & & & &&>

[0071] As shown at block 306 of FIG. 3, the CaBeSt-Net model was initially tested on two sets of ROIs and then incorporated into the analytical pipeline assessing TC in histologic WSIs with LUAD (right panel in FIG. 3). The pipeline also included StarDist, a DL model that provides a nuclear mask for the WSI. Overlaying the WSI nuclear mask onto the CA mask (outputted by the CaBeSt-Net model) identified the nuclei within the CA area. These nuclei were then counted, and TC was assessed as the ratio of malignant cell nuclei to all nuclei within the WSI or within an arbitrarily selected tumor area (TA).

[0072] Different sets of digital slides, as shown in FIG. 3 and Table 1, from archived formalin-fixed paraffin-embedded primary LUAD resections (1 unstained slide per case, 110 slides total) were assessed. Of these, 17 digital IHC stained slides were utilized for development of the Ep-Net model. To develop the CaBeSt-Net model, digital slides were prepared from restained tissues. In particular, unstained slides were first stained with H&E,144913-5900-09642065472-001009 WOPTthen digitized, destained and restained with IHC markers to visualize epithelium and digitized again. The IHC staining was carried in separate batches and executed through manual or autostainer-based protocols using antibodies against pancytokeratin (panCK) with or without thyroid transcription factor-1 (TTF1). In total, 39 H&E and IHC digital slide pairs were prepared for the CaBeSt-Net model development (Table 1). This model was tested on a set of ROIs from 4 digital H&E slides from CSMC and a separate set of ROIs from 20 digital H&E slides from the National Lung Cancer Screening Trial (NLST) [https: / / www.cancerimagingarchive.net / collection / nlst / ] repository with manual CA, BE and ST delineations provided by pathologist (BP), and subsequently on ROIs from 50 digital H&E and IHC slide pairs from restained tissues. The 50 digital H&E slides from the restained tissues were included in the study slides and used for TC assessments (Table 1). In this study, the cases had been diagnosed by experienced surgical pathologists as LU Ds, and TTF1 was used solely to improve the cancer mask for Al development.

[0073] At HLD, the IHC staining was carried out using the Ventana Benchmark Ultra autostainer (Roche Diagnostics, Indianapolis, IN). The staining was performed sequentially by first applying an anti-thyroid transcription factor 1 antibody (TTF1) (Roche, #790-4756) and the Ultra View universal DAB detection kit (Roche, #760-500) for visualization of epithelial cell nuclei. In the second step, a pan-cytokeratin (panCK) antibody cocktail (Roche, #760-2595) and the Ultra View detection kit with FastRed chromogen (Roche, #760-501) were applied to visualize keratins in the cytoplasm of the epithelial cells. Hematoxylin counterstain was performed last.

[0074] To prepare data for the Ep-Net model development, slides were reviewed at block 302 of FIG. 3. For example, the slides may be annotated to outline 372 ROIs. The ROIs were selected from areas containing well-preserved malignant or benign / normal and separately from areas lacking preserved epithelial cells (panels B, and C of FIG. 10, bottom row).

[0075] For CaBeSt-Net model development, as shown in block 304 of FIG. 3, 846 ROIs may be outlined in 39 IHC and H&E WSI pairs in the IHC WSIs. An ROI was labeled as (CA) if it contained cancer cells and stroma, as (BE) if it contained only benign epithelium or alveoli and stroma, or as (ST) if it contained stroma, bronchial cartilage, or necrosis in any proportion, and / or was devoid of preserved epithelial cells. Outlining and labeling of the 846 ROIs were performed using the ImageScope viewer and yielded 461 CA, 327 BE, and 58 ST ROIs. Other image viewing and annotation systems may also be used.

[0076] For CaBeSt-Net model testing, at block 306, 731 ROIs within the 50 IHC slides from the study set were identified. This set comprised 535 CA, 176 BE and 20 ST ROIs. In 154913-5900-09642065472-001009WOPTaddition, 183 and 110 test ROIs were extracted from the set of 4 (CSMC) and 20 (NLST, 5 - 6 ROIs / slide) H&E-stained slides and annotated to manually delineate and label the malignant epithelium (CA), benign epithelium (BE), and stroma (ST) within each ROI.

[0077] In the 50 H&E study set slides (right panel in FIG. 3, Table 1), tumor area with high density of tumor cells (TA) (1 TA per slide) was first circled using the ImageScope slide viewer and revised, in some configurations. The TC assessments in the TAs were compared with the assessments in corresponding entire WSIs.

[0078] The IHC ROIs from areas with well-preserved malignant or benign / normal epithelium were color deconvoluted, and the deconvoluted red dye image was thresholded. The resulting binary masks, with epithelial pixels equal to 1 and all other pixels equal to 0, were then smoothed and small holes (areas negative for panCK staining such as nuclei) were filled by mathematical morphology operators. Pixels in masks for the IHC ROIs from areas lacking well-preserved epithelial cells and / or containing necrosis (some of which show focal panCK staining) were set to 0. Finally, the IHC ROIs and corresponding masks were downsized to magnification equivalent to lOx and divided into non-overlapping 256 x 256-pixel tiles (N = 1,802). The tile-mask pairs were split randomly into training (N = 1,652) and validation (N = 150) sets.

[0079] To increase the Ep-Net model generalizability, the tile training set was augmented 9 times using random flipping, rotation and color modifications as described in. The Ep-Net model was trained for 20 epochs; the minibatch, initial learning rate, momentum, and L2-regularization for stochastic gradient descent optimizer were set to 16, 0.005, 0.9, and 0.001, respectively. The learning rate was halved every 2 epochs and reached 9.76E-6 at the end of training.

[0080] The CaBeSt-Net model was developed using ROIs from the 39 IHC and H&E slide pairs, as shown at block 304 of FIG. 3, through an automated approach in which low-resolution (5x magnification) IHC and H&E WSIs were converted to grayscale and then coarsely aligned by an image registration procedure.

[0081] FIG. 4A is a graphical illustration showing steps of a method to establish the Al model, according to aspects of the present disclosure. A workflow developing an CaBeSt-Net model predicting CA, BE, and ST masks in H&E ROIs is illustrated in FIG. 4. At block 402 of FIG. 4, ROI coordinates were then transferred from the IHC WSI to the H&E WSI using an intensity -based multimodal image registration. The matching IHC and H&E ROIs were paired, and then extracted from the full resolution WSIs. After extraction, the paired ROIs were finely164913-5900-09642065472-001009WOPTaligned by the same registration technique, and downsized (equivalent of lOx magnification). The multimodal image registration had been previously evaluated on ROIs from restained.

[0082] Block 402 of FIG. 4 shows coordinates of ROIs marked in the H4C WSI (with red marking epithelial cells) that were transferred to the H&E WSI through the low-resolution WSI registration technique. Block 404 of FIG. 4 shows corresponding H&E and IHC ROIs subsequently registered at full-resolution and the IHC ROIs were processed by the Ep-Net model to obtain epithelial masks (top and middle row). Pixels under the masks were labeled as BE (blue overlay) or CA (green overlay), and ST (no color) and together with the corresponding H&E ROIs divided into tiles (yellow grid) for augmentation and training of the CaBeSt-Net model.

[0083] Prior to the CaBeSt-Net model training, the tiles and corresponding masks were augmented 15 times through random perturbations of image coloration and geometry, and the obtained 187,056 H&E-mask tile pairs were used to train the CaBeSt-Net model. The training lasted 25 epochs; the minibatch, initial learning rate, momentum, and L2-regularization for stochastic gradient descent optimizer were set to 36, 0.0025, 0.9, and 0.001, respectively. The learning rate was halved every 2 epochs and reached 6.1E-7 at the end of training, in this example embodiment.

[0084] FIG. 4B is another graphical illustration showing steps of a method to establish the Al model, according to aspects of the present disclosure. At block 404 of FIG. 4B, the Ep-Net model was also applied to the IHC ROIs to obtain binary epithelial masks. Depending upon the label assigned, pixels in the epithelial mask with value of 1 (epithelium) were replaced with the value of 2 (for CA) or left unchanged (1 for BE). ST pixels retained 0 value. FIG. 4C is another graphical illustration showing steps of a method to establish the Al model, according to aspects of the present disclosure. Subsequently, the H&E ROIs paired with these masks (846 pairs) were divided into non-overlapping 256 x 256-pixel tiles (N = 12,306 mask-tile pairs) and split into training (N = 11,691 pairs) and validation (N = 615 pairs) sets, as shown in FIG. 4C.

[0085] The Ep-Net and CaBeSt-Net models were evaluated within the test time augmentation framework using global accuracy and degree of overlap metrics. Validations were based on measuring the global accuracy in respective held-out validation sets. Referring back to block 304 of FIG. 3, the CaBeSt-Net model was tested using three independent sets of ROIs comprising 183 (CSMC) and 110 (NLST) H&E ROIs had manually generated ground truth (GT) masks for CA, BE and ST areas. The other set comprised 731 H&E ROIs from the study slides with the GT masks for CA, BE, ST obtained from the paired IHC slides by the Ep-Net model. Below, global accuracy (gACC) reflecting the aggregated correctness of the CA,174913-5900-09642065472-001009WOPTBE, and ST pixel category predictions is provided. To quantify the degree of overlap between the GT and the prediction masks with CA, BE and ST pixels by the CaBeSt-Net model, the intersection over union (loU) for each predicted pixel category, and weighted loU (wIoU) for each ROI and for the ROI test sets were computed (Matlab ver. 2022a, function: evaluate S emanti c S egmentati on) .

[0086] The StarDist model was evaluated using ROIs with diverse composition and tumor cellularity from the internal test set (a) (CSMC, 10 ROIs) and the external test set (c) (NLST, 10 ROIs). In each ROI, nuclei detected by StarDist were inspected to identify undetected nuclei, removed objects that were not nuclei and corrected nuclear delineations (split or joint nuclei), providing ground truth nuclear masks for over 38,500 nuclei across these 20 ROIs. The nuclear detections and segmentations by StarDist were assessed by the average precision (AP), and gACC and wIoU metrics computed in each subset of the selected ROIs.

[0087] The evaluated CaBeSt-Net model was incorporated into the pipeline, and the pipeline was run to output a mask with CA, BE, and ST pixel predictions for each test ROI and each H&E study WSI. In a similar way, StarDist, a nuclear segmentation algorithm built into QuPath, was applied to obtain a mask of nuclei. The number of nuclei counted within the WSI assessed the total number of cells in the tissue section while the number of malignant cells was assessed through counting the nuclei within the CA mask. TC within the test ROI and WSI was assessed by dividing the number of malignant cells by the total number of cells. TC assessment within the TA (previously defined as part of the W SI) was obtained by restricting the malignant and other cell counts to TA only. A custom Matlab script was written to automate the TA and WSI-based TC assessments. Prior to running the pipeline, white background and ink mark pixels were excluded from the WSIs. TC assessments were conducted using the ImageScope viewer. Other viewers may also be used to conduct the assessments. In some embodiments, the system imported one H&E WSI with the TA contour to the viewer, assessed TC within the TA and in the WSI, and recorded both TC values in a spreadsheet. Based on TCs assessed in WSIs and TAs, tumor samples were evaluated for adequacy for molecular testing. A sample was considered inadequate if the assessed TC < 20%, a cut-point. TC < 10% and 30% cut-points were used for comparison. Using these cut-points, samples categorized as inadequate by each rater were counted for statistical analysis.

[0088] Descriptive statistics were applied to summarize TC evaluations derived from WSIs and TAs using mean values and standard deviations. The output of the Al models was quantitatively assessed by intraclass correlation coefficients (ICC), providing a measure of the reliability of continuous ratings across different raters. The ICC was applied to assess the 184913-5900-09642065472-001009WOPTreliability of TC ratings in ROIs and WSIs. Furthermore, a linear mixed effects model was utilized to explore the fixed effects of the area assessed (TAs vs. WSIs) on TC, and slide and rater as random effects to account for inherent variability. Additionally, a consistency of categorization into adequate versus inadequate samples for molecular testing among raters was evaluated using Fleiss' kappa, offering insight into the qualitative agreement beyond chance. All statistical tests were conducted as two-tailed with the alpha level of 0.05, unless otherwise specified.

[0089] For this study, the same slide sequential H&E and immunostaining was performed for 106 samples, collected 106 IHC-stained and 130 H&E-stained WSIs, annotated and extracted 2,242 ROIs, trained and evaluated Ep-Net and CaBeSt-Net models, and computationally assessed TC in 20 test ROIs and subsequently in 50 WSIs with LU AD using the developed pipeline. TC assessments by 3 pathologists and the pipeline were statistically evaluated and the samples were assessed for adequacy for molecular testing.

[0090] Referring to FIG. 5A, FIG. 5B, and FIG. 5C, examples of microscopically challenging H&E ROIs with manual GT from the NLST slide set and predictions by the CaBeSt-Net model are shown. In particular, FIG. 5 A shows tumor adjacent to stroma with inflammatory infiltrate and anthracotic pigments; FIG. 5B shows tumor cells in solid growth pattern in close proximity to inflammatory infiltrate; and FIG. 5C shows a benign gland adjacent to stroma with small vessels, immune cells and tumor glands with macrophages. Hatched arrows indicate pixels or cells missed by the CaBeSt-Net model. Solid arrows show example false positive CA pixel detections. The gACC and wIoU metrics reflecting model predictions in panels A-C are (93.57%; 0.883), (89.45%; 0.810), and (95.0%; 0.908), respectively.

[0091] Example H&E and IHC ROIs and prediction masks by the Ep-Net and CaBeSt-Net model in heterogeneous tumor areas are shown in FIG. 6A, FIG. 6B, FIG. 6C, FIG. The Ep-Net model was run on the IHC ROIs, and the CaBeSt-Net model was run on the H&E ROIs. Example H&E ROIs include areas that would be difficult to annotate manually in which exhibits poorly differentiated LU AD, shown in images 602 of FIG. 6A, tumor glands contain luminal macrophages, shown in images 604 of FIG. 6B, situations where the tumor is difficult to distinguish from stroma, shown in images 606 of FIG. 6C, and immune cells are adjacent to benign glands, shown in images 608 of FIG. 6D. Hatched arrows indicate pixels or cells missed by the CaBeSt-Net model. Solid arrows show example false positive CA pixel detections. The gACC and wIoU metrics reflecting CaBeSt-Net model prediction accuracy in panels A-D are (91.1%; 0.839), (91.6%; 0.847), (91.9%; 0.851), and (95.4%; 0.912), respectively. All IHC 194913-5900-09642065472-001009WOPTROIs are from slides restained with panCK+TTFl IHC. To better visualize prediction masks, the ROIs are shown at different scales.

[0092] Example masks outputted by the Al models for an automated TC assessment in a histology slide with LUAD are shown in FIG. 7A, FIG. 7B, and FIG. 7C. Specifically, image 702 shows an example H&E WSI; image 704, shows example predicted masks outputted by the CaBeSt-Net model; and image 706 shows example ROIs from the WSI with prediction masks and nuclear contours by the StarDist model overlaid.

[0093] As shown in the examples of FIGS. 5A-7C (also summarized in Table 2), an CaBeSt-Net model evaluated on H&E ROIs with manual-based or IHC-based GT each achieved high values of the gACC (> 91%) and wIoU (>0.83). By comparing the predicted and the IHC-based GT masks of CA, foci could be found in which the CA mask was thinner and / or isolated IHC stained cells were missed. False positive CA pixel detections were rare. The predicted BE masks were less thin than the corresponding GT masks (FIG. 5C, 608 in FIG.6D). Pixels missed in the CA masks can be seen in ROIs with the wIoU < 0.6 except in the ROI with the lowest wIoU due to poor image focus. Nevertheless, given that only 12.8% -15.5% of all CA pixels in test ROIs were missed and that the prediction masks of 98.3% of all test ROIs were characterized by wIoU > 0.628, the predicted masks yielded by the Al-pipeline can be considered very accurate.

[0094] When validated on respective hold-out sets, the Ep-Net and the CaBeSt-Net models achieved global accuracy of 98.09% and 90.34%, respectively. The observed accuracy of the Ep-Net model was considered sufficient to provide ground truth epithelial masks (IHC-based GT) for the 731 H&E ROIs from the study set in lieu of manual delineations for this set of ROIs.

[0095] The ROI test sets (a-c) (Table 1, Supplementary Table 1) were used to test the CaBeSt-Net model which achieved the gACC = 93.49% and wIoU = 0.880 (manual GT, CSMC slides), gACC = 91.28% and wIoU = 0.839 (IHC-based GT, CSMC slides), and gACC = 93.29% and wIoU = 0.877 (manual GT, NLST slides), respectively. The loU metrics for CA, BE and ST tissue categories are listed in Table 2.204913-5900-09642065472-001009WOPTTable 2. Performance metrics of the CaBeSt-Net model evaluated in test sets.accuracy metric(a) manual GT (CSMC) 9349 0.880 0.782 0.733 0 916 (b) IHC-based GT (CSMC) 91.28 0.839 0.772 0.822 0.874 (c) manual GT (NLST) 93.29 0.877 0.794 0.731 0.944Supplementary Table 1. ROI and pixel proportions in test sets used to evaluate the CaBeSt-Net model.test set (a) manual GT (b] IHC-based GT (c) manual GT (NLST) number of ROIs 183 731 110number of pixels 75,978,773 340,947,448 31,748,772% of ST pixels 7375% 6420% 71 84%% of BN pixels 2 57% 3.95% 5 54%% of CA pixels 2368% 31 84% 22.62%

[0096] Example CA, BE, and ST masks predicted by the CaBeSt-Net model in microscopically challenging H&E ROIs with dense and heterogenous cell populations from the test set with manual-based GT (FIG. 5) and from the test set with H4C -based GT (FIG. 6). FIG.12 shows example CA masks with low wIoU (wIoU < 0.628, comprising 1.66% of all 1024 wIoUs) in the combined test set. StarDist nuclei detection (AP > 0.93) and segmentation (gACC > 95%, wIoU > 0.91) accuracies in the ROIs with manual GT are shown in Supplementary Table 2. The confusion matrices, gACCs and wIoUs were computed using Matlab, ver. 2022a, function: evaluate SemanlicSegmenlalion. QI- first quartile, IQ -interquartile range of the wIoU distribution.Table 3. Intraclass correlation coefficients (ICC) assessing reliability of TC ratings in ROIs selected from the test set (a) (CSMC WSIs) and test set (c) (NLST WSIs).ICC [95% Cl] agreement10 ROIs, test set (a) 0 971091 0.991* perfect10 ROIs, test set (c) 0 991098 0 991* perfect- p < 00001Supplementary Table 2. Performance metrics of the StarDist model in selected test ROIs from test set (a) (CSMC WSIs) and test set (c) (NLST WSIs).4913-5900-09642065472-001009 WOPTaccuracy metrics10 ROIs test set (a) 955% 0.916 095110 ROIs test set (c) 975% 0.951 0932

[0097] The reliability of TC ratings by the Al in the 20 test ROIs was high (ICC > 0.97) suggesting perfect agreement between the raters (Table 3) and subsequent applicability of the pipeline to TC assessment in H&E WSIs.

[0098] Referring now to FIG. 8A, FIG. 8B, FIG. 8C, and FIG. 8D, graphs showing TC assessments and effect of TC assessment thresholding on sample selection for molecular testing by raters, according to aspects of the present disclosure. Discrepancies in TC assessments in WSIs shown in graph 802 and graph 804 TAs arbitrarily selected from within the WSIs are arranged by Al-based assessments in ascending order. Applying a cut-point (TC < 20%) to TC assessment of each rater provides the number of samples inadequate for molecular testing. The number of inadequate samples ranged from 11 to 32, as shown in graph 806 of FIG. 8C, and from 3 to 6, as shown in graph 808 of FIG. 8D. The number of samples deemed inadequate by Al was respectively 13 and 3. Supplementary Table 3 summarize the distributions of TC assessments.Supplementary Table 3. Summary table of TC assessments (Mean ± SD) in whole slide images and selected tumor areas._ tumor areas whole slide images rater support(TAs) [%](WSIs) [%] Pathologist-1 62.2 ± 22.0 363 ± 207 Pathologist-2 without Al support 53.6 ± 19.1 390 ± 17.0 Pathologist-3 41 4 ± 208 207 ± 144 Pathologist-1 47 5 ± 171 284 ± 136 Pathologist-2 with Al support 46.2 ± 16.2 304 ± 135 Pathologist-3 43.3 ± 15.2 273 ± 11 3Al-pipeline 44.4 ± 14.7 284 ± 127

[0099] Based on the linear modelling of fixed effects, the difference between the TA-based and WSI- based TC assessment of 19.3% was statistically significant (p < 0.001) after accounting for the random effects of H&E slides (variance attributed = 143.6%) and the rater (variance attributed = 69.9%). The agreement reflected by the ICC was moderate and statistically significant (p < 0.001, Table 4). ICC values remained approximately the same after224913-5900-09642065472-001009 WOPTAl was added as a 4th rater. In this analysis, a fixed set of up to 4 raters rating each slide with no generalization to a larger population of raters was assumed.Table 4. Intraclass correlation coefficients (ICC) assessing reliability of TC ratings in 50 H&E WSIs."<

[0100] The effect of the TC < 20% cut-point on the number of samples categorized as inadequate for molecular testing by all raters is shown in graphs 806 of FIG. 8C and 808 of FIG. 8D. Using this predefined cut-point, 5 samples (by WSI) and 2 samples (by TA) would be inadequate by consensus of all 3 pathologists. The same WSI and TA samples would be categorized as inadequate when Al is included as the 4th rater. The 5 samples categorized as inadequate using the WSI-based TC assessment include the 2 samples categorized as inadequate using the TA-based assessment. The lowest number of tumor cells counted in TAs by Al was 9,665 and the highest was 1,130,439 thereby far exceeding the minimum of 100 cells required for the testing.Table 5. Fleiss’ kappa values evaluating agreement between pathologists and Al qualifying samples for molecular testing (inadequate vs. adequate, TC < 20% cut-point) using WSI-based or TA-based TC assessments.TC assessment Fleiss’ kappa [95% Cl] agreement234913-5900-09642065472-001009WOPT

[0101] The reliability of agreement between the raters (pathologists and Al in selected combinations) qualifying samples for molecular testing using the 20% cut-point as reflected by Fleiss’ kappa is shown in Table 5. The agreement between Al and one of the pathologists, and that among all 3 pathologists was slight and accidental (p > 0.05), and moderate (p < 0.01) between Al and the other two pathologists when TC was assessed in WSIs. When TC was assessed in TAs, the agreement between Al and any pathologist, among all 3 pathologists, and all 3 pathologists with Al ranged from fair to substantial (p < 0.01) surpassing that expected by chance alone. Also, the agreement of all 3 pathologists with Al (p < 0.001) was higher than that among all 3 pathologists when using the TC < 10% and 30% cut-points in WSI and TA-based TC assessments (Supplementary Table 4). The alpha level was adjusted to 0.01 for each test in accordance with the Bonferroni correction, applied to account for the five tests conducted.244913-5900-09642065472-001009 WOPTSupplementary Table 4. Fleiss’ kappa values evaluating agreement between pathologists and Al qualifying samples for molecular testing (inadequate vs. adequate, 10% and 30% TC cutpoints) using WSI-based or TA-based TC assessments.""" & """ " ""&"" " " " "< " < " < "" <

[0102] Existing evidence suggests that low reproducibility of manual TC assessment in slides from LUAD can impact results of molecular testing. Although DL is applicable to routine diagnostic procedures in histopathology, the availability of reliable DL models that can assist 254913-5900-0964 2065472-001009 WOPTin TC assessment is limited. To address this need, a large volume of image data from IHC restained slides was generated, and an approach that can automatically mask cancer cells for counting and TC assessment was developed.

[0103] The observed percentages of CA pixels missed by our CaBeSt-Net model are much lower than the 30% cancer cell pixel miss rate reported previously. Moreover, to assess the cancer masking accuracy of the DL model in another study, ROIs positive and negative for tumor regions (as detected by their model) were inspected to assess the accuracy of cancer cell detection. In contrast to that approach, in the present study, the ratios of correctly classified pixels and area overlaps were measured within each test ROI with semantically established GT, thereby constituting a more in-depth and rigorous evaluation.

[0104] The performance metrics observed in the present study set ROIs with IHC -based GT (random location, approximately 15 ROIs / slide) suggest that the CaBeSt-Net model reliably distinguished CA areas from BE and ST areas in the 50 study slides used for TC assessment (Table 2, also shown in FIGS. 5 to 7). Given that these slides were stained in two separate batches, and each batch stained using different staining equipment and antibody cocktail, and scanned using different scanners, it may be inferred that this model is sufficiently robust to accommodate variabilities in slide preparation and image acquisition. It may also be inferred that the high CA masking performance demonstrated by the CaBeSt-Net model results from its capacity to learn tissue architecture from the H&E images through the IHC -based masks that more accurately locate epithelial cells and boundaries between these and other cell types than the manual GT masks. The efficacy of IHC -based model learning is confirmed by the high accuracy and loU metrics achieved in the H&E ROIs with manual GT (Table 2). These performance metrics and those reported by others suggest that the present model can predict reliable CA masks in H&E slides that are similar to the masks obtained through immunostaining.

[0105] A DeepLabV3+ architecture was selected to develop CaBeSt-Net, although other architectures may be possible.

[0106] In some example embodiments, Al pipelines in ROIs were validated in lieu of WSIs. In some studies, masks were qualitatively assessed to determine the predicted tumor and nuclear masks from selected ROIs, and then, based on the reliability of these masks adjusted the TC assessments in WSIs. In the present study, evaluation of the model’s masking performance in WSIs was also ROI-based. However, it was performed before the TCs were assessed, and instead a quantitative CA mask assessment with IHC -based reference was used.264913-5900-09642065472-001009WOPT

[0107] Incorporating the IHC -based GT from restained slides allowed for addressing the challenge in validating the model on WSIs, and produced a model applicable to confidently assess TC in WSIs. Among various advantages, the present proof-of-concept tool and approach may allow for the determination of TC more quickly, more accurately and more reproducibly. Visualization of the detected tumor mask and the TC value assessed are main benefits.

[0108] Assessment of TC is critical in determining a sample’s adequacy for molecular testing. Results obtained on the random effects by linear mixed effect modelling indicate that there is variability in TC assessment attributable to differences in slides and raters, with more variance attributed to the slides than to the raters. Moreover, the area selected (WSI vs. TA) has the statistically significant effect on TC assessment value (p < 0.001), supporting that TC assessments based on a TA within the WSI are higher than TC assessments based on the entire tissue visualized in the WSI. Therefore, in some embodiments, assessment may include selecting an area (TA) from within the WSI rather than the entire WSI for TC assessment followed by tissue macrodissection.

[0109] The consistency of categorizations into inadequate versus adequate samples for molecular testing (TC < 20%) among pathologists was moderate and statistically significant in TAs (p < 0.0001) (Table 5), while sample categorization based on TCs assessed in the WSIs was slight (p = 0.106). However, inclusion of Al approach, according to the present disclosure, as an additional rater improved the consistency of sample categorization (increased Fleiss’ kappa values and improvements in statistical significance were observed) both in the WSIs and TAs. The agreement between each of the three pathologists and Al in sample categorization based on TC in the TAs was statistically significant (p < 0.01, the Bonferroni adjusted alpha level for 5 tests). This was not the case in the WSI-based categorization due to lack of agreement between Al and one of the pathologists. The agreements in TAs were consistently higher than the respective agreements in WSIs and the consistency and statistical significance held despite various cut-points.

[0110] Lower agreements and lack of agreement in one case observed in WSIs may have been associated with increased difficulty in assessing TC in larger and more heterogenous areas in the sample, and hence increased discrepancies in sample categorization by pathologists. Hence, pathologists might consider using the tumor cell mask and proportion of tumor cells within the sample (WSIs or TAs) provided by Al to objectify their sample categorization.[oni] The CaBeSt-Net model training set included ROIs with two-category (CA and ST, BE and ST) and one-category (ST) GT masks. In some aspects, introducing the BE as an additional category in the training may increase the model’s discriminatory capability. To test 274913-5900-09642065472-001009WOPTwhether the CaBeSt-Net model learned features discriminating adjacent benign and cancer cells, ROIs with the manual GT were utilized. Although the 4 H&E CSMC and 20 NLST slides available for manual annotations contained only 11 ROIs with three-category GTs containing benign and cancer cells in close proximity, improvements may be achieved by blending ROIs using generative Al models.

[0112] The present processing pipeline was trained using data from two institutions, but due to availability of LU AD slides restained with H4C at HLD, testing was conducted on slides from CSMC only. To address this, test slides at CSMC were prepared in two separate batches simulating different slide staining and scanning conditions. Also, the present pipeline was tested on slides from resected tumors. Although TC is frequently assessed in needle core biopsies, collecting biopsies, and performing TC assessment in biopsy slides was beyond the scope of this study.

[0113] Example ROIs from digital slides with restained tissues acquired at CSMC and HLD and used herein are shown in FIG. 9A, FIG. 9B, and FIG. 9C. Specifically, images 902 of FIG. 9A shows IHC ROI exhibiting a light brown / red precipitate throughout nuclei and cytoplasm of cancer cells. Images 904 of FIG. 9B shows IHC ROI with cytoplasm of cancer cells stained crisp red. Images 906 of FIG. 9C shows IHC ROI with purple precipitate in nuclei and cytoplasm. All slides were counterstained with hematoxylin. Other counterstains may also be used.

[0114] At CSMC, the IHC staining was carried out in two batches: batch 1 was stained with panCK alone to visualize keratins in the cytoplasm of epithelial cells (N = 40, Table 1) while batch 2 was stained with panCK+TTFl cocktail to visualize both the cytoplasm and the nuclei of epithelial cells (N = 34, Table 1). Application of the panCK primary antibody was followed by application of the secondary antibody. ImPRESS anti-mouse alkaline phosphatase and detection of keratins by the red dye was also performed.

[0115] Then, the slides were counterstained with hematoxylin. In some example embodiments, this includes automatically staining the slides. For example, some slides may be subsequently stained automatically using an automatic platform. Initial steps involved heat-induced antigen retrieval with cell conditioning solution. Slides were then incubated in the prediluted rabbit nuclear anti-TTFl primary antibody for 32 minutes at 37°C, followed by the application of anti-rabbit NP and anti-NP AP multimer detection systems. Subsequently, a prediluted mouse anti-panCK cocktail was applied for 32 minutes at 37°C, and anti-mouse NP and anti-NP AP multimer detection systems were applied. Finally, the Discovery Red Kit was applied to provide a purple precipitate marking expression of TTF1 and panCK. Mayer’s 284913-5900-09642065472-001009WOPThematoxylin, or another hematoxylin may be used as the counterstain. Prior to scanning, the slides were air dried and mounted. Example ROIs from both Batch 1 and 2 are shown in images 904 of FIG. 9B and images 906 of FIG. 9C.

[0116] An example IHC stained slide and ROIs box-outlined that were used to train the epithelium masking model (Ep-Net) is shown in FIG. 10A, FIG. 10B, and FIG. 10C. In particular, image 1002 of FIG. 10A shows a whole slide image and consecutively numbered ROIs with distinct histology and staining intensity. Images 1004 of FIG. 10B shows ROIs with well-preserved epithelium (top row: malignant, middle row: mucinous glands) and epithelial masks (green overlay). Images 1006 of FIG. 10C shows ROIs from areas lacking preserved epithelial cells and / or containing necrosis (some of which showing focal panCK staining) with all pixels in the epithelial mask set to 0 (no green overlay).

[0117] To demonstrate the performance of the CaBeSt-Net model at predicting pixels of malignant cell areas (CA), benign / normal epithelium (BE) and stroma (ST) in test ROIs, confusion matrices and boxplots of accuracy metrics were prepared, as shown in FIG. 11 A and FIG. 11B. In particular, graph 1102 of FIG. 11 A, shows: confusion matrices with % correct pixel predictions for test set with manual GT (a), graph 1104 shows confusion matrices with % correct pixel predictions for test set IHC -based GT, and graph 1106 shows confusion matrices with % correct pixel predictions for test set using manual GT (c) (NLST WSIs), with shading reflecting the % value (from 0% -white to 100% - black). Graph 1108 of FIG. 11B shows boxplots visualizing gACC and wIoU distributions in the combined test set comprising 1024 ROIs. wIoU distribution outliers (values lower than (QI - 1.5*IQ) = 0.628) are indicated by “x.”

[0118] Example H&E ROIs with predicted masks and the corresponding wIoU < 0.628 are shown in FIG. 12 A, FIG. 12B, FIG. 12C, FIG, 12D, FIG. 12E, FIG. 12F, and FIG. 12G.. In particular, FIG. 12A shows results from the CaBeSt-Net model characterized by a low wIoU. The gACC and wIoU metrics for the predicted masks are as follows: graphs 1202 of FIG. 12A show gACC: 68.9%; wlol: 0.483), graphs 1204 of FIG. 12B show gACC: 71.2%; wIoU: 0.5260, graphs 1206 of FIG. 12C show gACC: 69.9%; wIoU: 0.533, graphs 1208 of FIG. 12D show gACC: 73.2%; wIoU: 0.559, graphs 1210 of FIG. 12E show gACC: 72.2%; wIoU: 0.559, graphs 1212 of FIG. 12F show gACC: 73.9%; wIoU: 0.568, and graphs 1214 of FIG. 12G show gACC: 75.5%; wIoU: 0.592. These wIoUs are outliers in the wIoU distribution shown in FIG.11.

[0119] As demonstrated herein, histologic slides restained with immunostains expressed in epithelial cells can overcome challenges in generating fine and objective annotations and 294913-5900-09642065472-001009WOPTimprove development of Al models applicable to masking cancer, benign and stromal areas in digital H&E slides. By visualizing masks and measuring proportions of cancer cells to other cells in whole slide images of lung adenocarcinoma, the proposed Al pipeline can assist in assessing tumor cellularity for molecular testing.

[0120] As described herein, the present Al-based approach may be used in a wide variety of applications, including applications for determining TC and providing a visual map of detected cancer from image data. While various examples of using trained Al models to recognize immune cells in histologic slides using immunohistochemistry are described herein, the present approach may also be scalable and have broader application. For instance, the present approach may be extended to various markings, as well as recognition of a wide variety of other cell and tissue types.ALTERNATIVE IMPLEMENTATIONS

[0121] Alternative implementation 1. A computer-implemented method for determining tumor cellularity in a tissue sample, the method comprising: obtaining, by at least one processor, and via a communications interface, imaging data associated with a tissue sample; predicting, by the at least one processor, and based on an application of a first trained artificial intelligence (Al) model configured to predict at least one target tissue in a region of interest (ROI) of the tissue sample to the imaging data, a predicted target tissue within the ROI of the tissue sample; predicting, by the at least one processor, and based on an application of a second trained Al model to the imaging data, a predicted cellular component count of at least one cellular component within the predicted target tissue; and determining, by the at least one processor, and based on the predicted cellular component count, data indicative of a tumor cellularity in the tissue sample.

[0122] Alternative implementation 2. The computer-implemented method of implementation 1, wherein the first trained Al model is further configured to predict at least one tumor tissue in the predicted target tissue.

[0123] Alternative implementation 3. The computer-implemented method of implementation 1, wherein determining the data indicative of a tumor cellularity in the tissue sample includes determining, by the at least one processor, the data indicative of a tumor cellularity by dividing a count of nuclei associated with pixels of the imaging data classified as malignant epithelium by a total count of nuclei within the ROI or in the tissue sample.

[0124] Alternative implementation 4. The computer-implemented method of implementation 1, wherein the method further comprises: establishing, by the at least one304913-5900-09642065472-001009WOPTprocessor, image pixels corresponding to the predicted target tissue; and generating, by the at least one processor, at least one tissue mask based at least on the image pixels.

[0125] Alternative implementation 5. The computer-implemented method of implementation 4, wherein the tissue mask comprises pixel-wise classifications of (i) malignant epithelium, (ii) benign epithelium, (iii) stroma, or (iv) any combination of (i)-(iii).

[0126] Alternative implementation 6. The computer-implemented method of implementation 4, wherein the predicting of the predicted cellular component count is based on an application of the second trained Al model to the at least one tissue mask.

[0127] Alternative implementation 7. The computer-implemented method of implementation 4, wherein the at least one cellular component includes a cell nucleus.

[0128] Alternative implementation 8. The computer-implemented method of implementation 4, wherein the method further comprises predicting a total count of cells in (i) the ROI, (ii), the tissue sample, or (iii) any combination of (i)-(ii) based on an application of the second Al model to the imaging data.

[0129] Alternative implementation 9. The computer-implemented method of implementation 1, further comprising selecting a predicted tumor area within the image characterized by the imaging data, and computing a first tumor cellularity within the tumor area and a second tumor cellularity within an entire image characterized by the imaging data.

[0130] Alternative implementation 10. The computer-implemented method of implementation 1, wherein the method further comprises generating, by the at least one processor, and based on the data indicative of the tumor cellularity, a report indicative of the tumor cellularity in the tissue sample.

[0131] Alternative implementation 11. The computer-implemented method of implementation 10, wherein generating the report indicative of the tumor cellularity includes generating, based on an application of a third trained Al model to at least the data indicative of a tumor cellularity of the tissue sample, a treatment recommendation.

[0132] Alternative implementation 12. The computer-implemented method of implementation 11, further comprising transmitting, by the at least one processor, and via the communications interface, a signal to a treatment apparatus that causes the treatment apparatus to (i) initiate a treatment, (ii) modify a treatment, (iii), pause a treatment, (iv) cease a treatment, or (v) any combination of (i)-(iv) based at least in part on the determined tumor cellularity.

[0133] Alternative implementation 13. The computer-implemented method of implementation 1, wherein the first trained Al model or the second trained Al model includes (i) a decision tree, (ii) a Bayesian network, (iii) an artificial neural network, (iv) a support 314913-5900-09642065472-001009WOPTvector machine, (v) a convolutional neural network, (vi) a capsule network, or (vii) any combination of (i)-(vi).

[0134] Alternative implementation 14. The computer-implemented method of implementation 1, further comprising: processing, by the at least one processor, the imaging data to generate data indicative of the imaging data; and generating an input dataset for respective first and second input layers of the first and the second trained Al models based on the data indicative of the imaging data.

[0135] Alternative implementation 15. The computer-implemented method of implementation 1, wherein the tissue sample includes a biopsy specimen, and wherein the imaging data is associated with the biopsy specimen.

[0136] Alternative implementation 16. A system for determining tumor cellularity in a tissue sample, the system comprising: a communications interface configured to obtain imaging data associated with a tissue sample; a storage device configured to store the imaging data; a memory storing instructions; and at least one processor communicatively coupled with the memory and configured to execute the instructions to: predict, based on an application of a first trained artificial intelligence (Al) model configured to predict at least one target tissue in a region of interest (ROI) of the tissue sample to the imaging data, a predicted target tissue within the ROI of the tissue sample; predict, based on an application of a second trained Al model to the imaging data, a predicted cellular component count of at least one cellular component within the predicted target tissue; and determine, based on the predicted cellular component count, data indicative of a tumor cellularity in the tissue sample.

[0137] Alternative implementation 17. The system of implementation 16, wherein the first trained Al model is further configured to predict at least one tumor tissue in the target tissue.

[0138] Alternative implementation 18. The system of implementation 16, wherein the at least one processor is further configured to: establish image pixels corresponding to the predicted target tissue; and generate at least one tissue mask based at least on the image pixels.

[0139] Alternative implementation 19. The system of implementation 18, wherein the tissue mask comprises pixel-wise classifications of (i) malignant epithelium, (ii) benign epithelium, (iii) stroma, or (iv) any combination of (i)-(iii).

[0140] Alternative implementation 20. The system of implementation 18, wherein the predicting of the predicted cellular component count is based on an application of the second trained Al model to the at least one tissue mask.

[0141] Alternative implementation 21. The system of implementation 16, wherein the at least one cellular component includes a cell nucleus.324913-5900-09642065472-001009WOPT

[0142] Alternative implementation 22. The system of implementation 16, wherein the at least one processor is further configured to predict a total count of cells in (i) the ROI, (ii), the tissue sample, or (iii) any combination of (i)-(ii) based on an application of the second Al model to the imaging data.

[0143] Alternative implementation 23. The system of implementation 16, wherein the at least one processor is further configured to generate, by the at least one processor, and based on the data indicative of the tumor cellularity, a report indicative of the tumor cellularity in the tissue sample.

[0144] Alternative implementation 24. The system of implementation 16, wherein the first trained Al model or the second trained Al model includes (i) a decision tree, (ii) a Bayesian network, (iii) an artificial neural network, (iv) a support vector machine, (v) a convolutional neural network, (vi) a capsule network, or (vii) any combination of (i)-(vi).

[0145] Alternative implementation 25. The system of implementation 16, wherein the tissue sample includes a biopsy specimen, and wherein the imaging data is associated with the biopsy specimen.

[0146] One or more elements or aspects or steps, or any portion(s) thereof, from one or more of any of the claims below can be combined with one or more elements or aspects or steps, or any portion(s) thereof, from one or more of any of the other claims or combinations thereof, to form one or more additional implementations and / or claims of the present disclosure.

[0147] While the present disclosure has been described with reference to one or more embodiments or implementations, those skilled in the art will recognize that many changes may be made thereto without departing from the spirit and scope of the present disclosure. Each of these implementations and variations thereof is contemplated as falling within the spirit and scope of the present disclosure. It is also contemplated that additional implementations according to aspects of the present disclosure may combine any number of features from any of the implementations described herein.334913-5900-09642065472-001009 WOPT

Claims

CLAIMSWhat is claimed is:

1. A computer-implemented method for determining tumor cellularity in a tissue sample, the method comprising:obtaining, by at least one processor, and via a communications interface, imaging data associated with a tissue sample;predicting, by the at least one processor, and based on an application of a first trained artificial intelligence (Al) model configured to predict at least one target tissue in a region of interest (RO I) of the tissue sample to the imaging data, a predicted target tissue within the ROI of the tissue sample;predicting, by the at least one processor, and based on an application of a second trained Al model to the imaging data, a predicted cellular component count of at least one cellular component within the predicted target tissue; and determining, by the at least one processor, and based on the predicted cellular component count, data indicative of a tumor cellularity in the tissue sample.

2. The computer-implemented method of claim 1, wherein the first trained Al model is further configured to predict at least one tumor tissue in the predicted target tissue.

3. The computer-implemented method of claim 1 , wherein determining the data indicative of a tumor cellularity in the tissue sample includes determining, by the at least one processor, the data indicative of a tumor cellularity by dividing a count of nuclei associated with pixels of the imaging data classified as malignant epithelium by a total count of nuclei within the ROI or in the tissue sample.

4. The computer-implemented method of claim 1, wherein the method further comprises:establishing, by the at least one processor, image pixels corresponding to the predicted target tissue; andgenerating, by the at least one processor, at least one tissue mask based at least on the image pixels.

5. The computer-implemented method of claim 4, wherein the tissue mask comprises pixel-wise classifications of (i) malignant epithelium, (ii) benign epithelium, (iii) stroma, or (iv) any combination of (i)-(iii).344913-5900-09642065472-001009 WOPT6. The computer-implemented method of claim 4, wherein the predicting of the predicted cellular component count is based on an application of the second trained Al model to the at least one tissue mask.

7. The computer-implemented method of claim 4, wherein the at least one cellular component includes a cell nucleus.

8. The computer-implemented method of claim 4, wherein the method further comprises predicting a total count of cells in (i) the ROI, (ii), the tissue sample, or (iii) any combination of (i)-(ii) based on an application of the second Al model to the imaging data.

9. The computer-implemented method of claim 1, further comprising selecting a predicted tumor area within the tissue sample, and computing a first tumor cellularity within the tumor area and a second tumor cellularity within an entire image characterized by the imaging data.

10. The computer-implemented method of claim 1, wherein the method further comprises generating, by the at least one processor, and based on the data indicative of the tumor cellularity, a report indicative of the tumor cellularity in the tissue sample.

11. The computer-implemented method of claim 10, wherein generating the report indicative of the tumor cellularity includes generating, based on an application of a third trained Al model to at least the data indicative of a tumor cellularity of the tissue sample, a treatment recommendation.

12. The computer-implemented method of claim 11, further comprising transmitting, by the at least one processor, and via the communications interface, a signal to a treatment apparatus that causes the treatment apparatus to (i) initiate a treatment, (ii) modify a treatment, (iii), pause a treatment, (iv) cease a treatment, or (v) any combination of (i)-(iv) based at least in part on the determined tumor cellularity.

13. The computer-implemented method of claim 1, wherein the first trained Al model or the second trained Al model includes (i) a decision tree, (ii) a Bayesian network, (iii) an354913-5900-09642065472-001009 WOPTartificial neural network, (iv) a support vector machine, (v) a convolutional neural network, (vi) a capsule network, or (vii) any combination of (i)-(vi).

14. The computer-implemented method of claim 1, further comprising:processing, by the at least one processor, the imaging data to generate data indicative of the imaging data; andgenerating an input dataset for respective first and second input layers of the first and the second trained Al models based on the data indicative of the imaging data.

15. The computer-implemented method of claim 1, wherein the tissue sample includes a biopsy specimen, and wherein the imaging data is associated with the biopsy specimen.

16. A system for determining tumor cellularity in a tissue sample, the system comprising:a communications interface configured to obtain imaging data associated with a tissue sample;a storage device configured to store the imaging data;a memory storing instructions; andat least one processor communicatively coupled with the memory and configured to execute the instructions to:predict, based on an application of a first trained artificial intelligence (Al) model configured to predict at least one target tissue in a region of interest (RO I) of the tissue sample to the imaging data, a predicted target tissue within the ROI of the tissue sample;predict, based on an application of a second trained Al model to the imaging data, a predicted cellular component count of at least one cellular component within the predicted target tissue; and determine, based on the predicted cellular component count, data indicative of a tumor cellularity in the tissue sample.

17. The system of claim 16, wherein the first trained Al model is further configured to predict at least one tumor tissue in the target tissue.

18. The system of claim 16, wherein the at least one processor is further configured to: establish image pixels corresponding to the predicted target tissue; and364913-5900-09642065472-001009 WOPTgenerate at least one tissue mask based at least on the image pixels.

19. The system of claim 18, wherein the tissue mask comprises pixel -wise classifications of (i) malignant epithelium, (ii) benign epithelium, (iii) stroma, or (iv) any combination of (i)- (iii).

20. The system of claim 18, wherein the predicting of the predicted cellular component count is based on an application of the second trained Al model to the at least one tissue mask.

21. The system of claim 16, wherein the at least one cellular component includes a cell nucleus.

22. The system of claim 16, wherein the at least one processor is further configured to predict a total count of cells in (i) the ROI, (ii), the tissue sample, or (iii) any combination of (i)-(ii) based on an application of the second Al model to the imaging data.

23. The system of claim 16, wherein the at least one processor is further configured to generate, by the at least one processor, and based on the data indicative of the tumor cellularity, a report indicative of the tumor cellularity in the tissue sample.

24. The system of claim 16, wherein the first trained Al model or the second trained Al model includes (i) a decision tree, (ii) a Bayesian network, (iii) an artificial neural network, (iv) a support vector machine, (v) a convolutional neural network, (vi) a capsule network, or (vii) any combination of (i)-(vi).

25. The system of claim 16, wherein the tissue sample includes a biopsy specimen, and wherein the imaging data is associated with the biopsy specimen.374913-5900-09642065472-001009 WOPT

Images