WO2025199073A1 - Generating medical prognosis using glandular and immune architecture features - Google Patents

Abstract

The present disclosure, in some embodiments, relates to a method that includes accessing digitized pathology imaging data of tissue excised from a cancer patient. The digitized pathology imaging data has been segmented to identify one or more of glandular regions, immune cells, and invasive cribriform adenocarcinoma (ICC). A plurality of features are extracted from the digitized pathology imaging data. The plurality of features include one or more of immune cell spatial features, glandular morphometric features, and ICC area features. The plurality of features are provided to a machine learning model that is trained to generate a medical prognosis for the cancer patient using the plurality of features.

Description

GENERATING MEDICAL PROGNOSIS USING GLANDULAR AND IMMUNE ARCHITECTURE FEATURES

REFERENCE TO RELATED APPLICATION

[0001] This Application claims the benefit of U.S. Provisional Application No. 63/568,707, filed on March 22, 2024, the contents of which are incorporated by reference in their entirety.

FEDERAL FUNDING INFORMATION

[0002] This invention was made with government support under R01CA268287 awarded by the National Institutes of Health/National Cancer Institute. The government has certain rights in the invention.

BACKGROUND

[0003] Prostate cancer is an uncontrolled growth of cells in the prostate, a small walnut-shaped gland in males that produces seminal fluid that nourishes and transports sperm. Prostate cancer is one of the most common types of cancer. Prostate cancer typically causes no symptoms in its early stages. While some types of prostate cancer grow slowly and may need minimal or even no treatment, other types are aggressive and can spread quickly. Prostate cancer that's detected when it's still confined to the prostate gland has the best chance for successful treatment.

BRIEF DESCRIPTION OF THE DRAWINGS

[0004] The accompanying drawings, which are incorporated in and constitute a part of the specification, illustrate various example operations, apparatus, methods, and other example embodiments of various aspects discussed herein. It will be appreciated that the illustrated element boundaries (e.g., boxes, groups of boxes, or other shapes) in the figures represent one example of the boundaries. One of ordinary skill in the art will appreciate that, in some examples, one element can be designed as multiple elements or that multiple elements can be designed as one element. In some examples, an element shown as an internal component of another element may be implemented as an external component and vice versa. Furthermore, elements may not be drawn to scale. [0005] Fig. 1 illustrates some embodiments of a block diagram of an assessment system configured to make a medical prognosis for a cancer patient.

[0006] Fig. 2 illustrates a flow diagram showing some embodiments of a method of making a medical prognosis relating to a cancer patient.

[0007] Fig. 3 illustrates some embodiments of a biochemical recurrence (BCR) assessment system configured to make a medical prognosis relating to BCR in a cancer patient.

[0008] Fig. 4 illustrates exemplary Kaplan Meier survival curves corresponding to a disclosed BCR assessment system trained to generate a medical prognosis solely using cribriform area index.

[0009] Fig. 5A illustrates exemplary Kaplan Meier survival curves corresponding to a disclosed BCR assessment system trained to generate a medical prognosis solely using cribriform area index and glandular morphometric features.

[00010] Fig. 5B illustrates a graph showing an exemplary receiver operating characteristic (ROC) curve for the disclosed BCR assessment system trained to generate a medical prognosis using cribriform area index and glandular morphometric features.

[00011] Fig. 6 illustrates exemplary Kaplan Meier survival curves corresponding to a disclosed BCR assessment system trained to generate a medical prognosis using CAI, glandular morphometric features, and immune cell spatial features.

[00012] Fig. 7 illustrates some additional embodiments of a block diagram of a BCR assessment system configured to make a medical prognosis relating to BCR in a cancer patient.

[00013] Fig. 8 illustrates some embodiments of a block diagram of an apparatus configured to make a medical prognosis relating to a cancer patient.

DETAILED DESCRIPTION

[00014] The description herein is made with reference to the drawings, wherein like reference numerals are generally utilized to refer to like elements throughout, and wherein the various structures are not necessarily drawn to scale. In the following description, for purposes of explanation, numerous specific details are set forth in order to facilitate understanding. It may be evident, however, to one of ordinary skill in the art, that one or more aspects described herein may be practiced with a lesser degree of these specific details. In other instances, known structures and devices are shown in block diagram form to facilitate understanding.

[00015] Prostate cancer is often detected using screening tests. For example, blood tests may be used to detect levels of prostate-specific antigen (PSA), which indicate unusual growth of prostate tissue. If high levels of PSA are detected, a diagnosis of prostate cancer typically requires a biopsy that removes tissue from a prostate. Once a patient has been diagnosed with prostate cancer, treatment options for the patient may vary depending on factors such as cancer stage, aggressiveness, overall patient health, and/or the like. If prostate cancer is determined to be aggressive, a radical prostatectomy may be performed to treat the prostate cancer.

[00016] During a radical prostatectomy, an entire prostate gland, some surrounding tissue, and the seminal vesicles may be removed. The goal of a radical prostatectomy is to remove all cancer cells. In some cases, after a radical prostatectomy, PSA levels may begin to rise again in a patient’s blood. Biochemical recurrence (BCR) is a condition in which PSA levels in the blood of a prostate cancer patient increase after treatment with surgery and/or radiation (e.g., to above levels exceeding 0.2 ng/mL). Men with prostate cancer that experience BCR after a radical prostatectomy are at elevated risk for metastasis and prostate cancer specific mortality. Therefore, the early identification of a risk of BCR can help identify men who might benefit from adjuvant therapies like radiation, docetaxel, and/or the like.

[00017] The present disclosure relates to a method and apparatus configured to generate a medical prognosis that identifies cancer patients at a high-risk for biochemical recurrence utilizing an artificial intelligence-based computational pathology model which involves extracting features relating to a glandular morphology, an invasive cribriform adenocarcinoma (ICC) area, and/or a spatial architecture of immune cells (e.g., tumor-infiltrating lymphocytes) from digitized pathology imaging data. In some embodiments, the method may be performed by accessing digitized pathology imaging data including one or more digitized pathology images of prostate tissue from a cancer patient having undergone a radical prostatectomy. The digitized pathology imaging data may be segmented to identify one or more of glandular regions, immune cells, and ICC. A plurality of features are extracted from the digitized pathology imaging data. The plurality of features include one or more of immune cell spatial features, glandular morphometric features, and ICC area features. The plurality of features are provided to a machine learning model that is trained to generate a medical prognosis regarding a risk of biochemical recurrence (BCR) for the cancer patient. It has been determined that features extracted from glandular tissue, ICC, and immune cells have a strong correlation with BCR. Thus, such features are able to be used by a machine learning classifier to generate an accurate prognosis of BCR, which can be used by health care professionals to make a more informed decision relating to the treatment of a cancer patient.

[00018] Fig. 1 illustrates some embodiments of a block diagram of an assessment system 100 configured to make a medical prognosis for a cancer patient.

[00019] The assessment system 100 comprises a memory 101 configured to store digitized pathology imaging data 102. The digitized pathology imaging data 102 includes imaging data from a cancer patient (e.g., a prostate cancer patient, a breast cancer patient, a cervical cancer patient, and/or the like). In some embodiments, the digitized pathology imaging data 102 comprises one or more digitized pathology images 104 (e.g., one or more digitized biopsy slides) obtained from a pathological tissue sample taken from the cancer patient and stored in the memory 101 . In some embodiments, the one or more digitized pathology images 104 may include glandular tissue. For example, the one or more digitized pathology images 104 may comprise images of prostate tissue removed from the cancer patient during a radical prostatectomy.

[00020] In some embodiments, a segmentation tool 106 may be configured to access the digitized pathology imaging data 102. The segmentation tool 106 is further configured to identify one or more of glandular regions 114, invasive cribriform adenocarcinoma (ICC) 116, and immune cells 118 within the digitized pathology imaging data 102 (e.g., within the one or more digitized pathology images 104). In some embodiments, the segmentation tool 106 comprises a first segmentation stage 108 configured to segment the one or more digitized pathology images 104 to identify one or more tumor regions 110. The segmentation tool 106 further comprises a second segmentation stage 112 that is downstream of the first segmentation stage 108. The second segmentation stage 112 is configured to segment the one or more tumor regions 110 to identify one or more of the glandular regions 114, the ICC 116, and the immune cells 118 within the tumor regions 110. In some embodiments, the immune cells 118 may comprise tumor infiltrating lymphocytes (TILs). In some embodiments, the segmented digitized pathology images may be saved in the memory 101 as part of the digitized pathology imaging data 102.

[00021] A feature extraction tool 120 is configured to extract a plurality of features 121 from the digitized pathology imaging data 102. In some embodiments, the plurality of features 121 may comprise one or more of immune cell spatial features 122, glandular morphometric features 124, and ICC area features 126. In some embodiments, the immune cell spatial features 122 may comprise features that describe an architecture of TIL nuclei and/or non-TIL nuclei (e.g., that characterize spatial arrangements and/or interactions between TIL nuclei and/or non-TIL nuclei) within the tumor region 110.

[00022] The plurality of features 121 are provided to a machine learning model 128 that has been trained to generate a medical prognosis 130. In some embodiments, the medical prognosis 130 may relate to a survival e.g., an overall survival, a disease free survival, etc.), a treatment response, and/or the like of the cancer patient. In some embodiments, the medical prognosis 130 may correspond to a determination as to whether or not the cancer patient will experience biochemical recurrence (BCR) free survival and/or a timeframe of BCR. For example, the medical prognosis 130 may categorize the cancer patient as having a low-risk 132 of BCR or a high-risk 134 of BCR. While post-prostatectomy treatment is effective in reducing metastasis and disease-specific death, it is not universally suitable due to the low overall mortality rate of prostate cancer. Therefore, by providing an accurate risk estimate for BCR post-surgery, the medical prognosis 130 can assist in identifying cancer patients that have a high likelihood of BCR and that may benefit from adjuvant therapy (e.g., like radiation, docetaxel, and/or the like), while avoiding unnecessary treatment for those at low-risk of BCR.

[00023] Fig. 2 illustrates a flow diagram showing some embodiments of a method 200 of making a medical prognosis relating to a cancer patient.

[00024] While the disclosed method 200 is illustrated and described herein as a series of acts or events, it will be appreciated that the illustrated ordering of such acts or events are not to be interpreted in a limiting sense. For example, some acts may occur in different orders and/or concurrently with other acts or events apart from those illustrated and/or described herein. In addition, not all illustrated acts may be required to implement one or more aspects or embodiments of the description herein. Further, one or more of the acts depicted herein may be carried out in one or more separate acts and/or phases.

[00025] At act 202, digitized pathology imaging data from a cancer patient is accessed. In some embodiments, the digitized pathology imaging data may comprise images of digitized pathology slides of a prostate tissue sample removed from the cancer patient during a radical prostatectomy.

[00026] At act 204, the digitized pathology imaging data may be segmented to identify one or more of glandular regions (e.g., glandular lumina), invasive cribriform adenocarcinoma (ICC), and immune cells (e.g., TIL). In some embodiments, the digitized pathology imaging data may be segmented according to acts 206-208.

[00027] At act 206, a first segmentation process is performed on the digitized pathology imaging data to identify a tumor region within the digitized pathology imaging data.

[00028] At act 208, a second segmentation process is performed on the digitized pathology imaging data to identify one or more of glandular lumina, immune cells, and tumor-infiltrating lymphocytes (TIL) within the tumor region.

[00029] At act 210, a plurality of features are extracted from the digitized pathology imaging data. In some embodiments, the plurality of features may be extracted according to one or more of acts 212-216.

[00030] At act 212, one or more immune cell spatial features are extracted using the immune cells identified in the digitized pathology imaging data. In some embodiments, the immune cell spatial features may comprise features that describe an architecture of TIL nuclei and/or non-TIL nuclei within the tumor region.

[00031] At act 214, one or more glandular morphometric features are extracted using the glandular lumina identified in the digitized pathology imaging data.

[00032] At act 216, one or more ICC area features are generated using the ICC identified in the digitized pathology imaging data.

[00033] At act 218, a machine learning model is operated to generate a medical prognosis using one or more of the immune cell spatial features, the glandular morphometric features, and the ICC area features. In some embodiments, the medical prognosis relates to a risk of biochemical recurrence (BCR) for the cancer patient.

[00034] At act 220, a treatment may be provided to the cancer patient based upon the medical prognosis, in some embodiments. For example, based upon the medical prognosis it may be determined that postoperative treatment may be beneficial to a patient and post operative treatment (e.g., such as radiation, docetaxel, and/or the like) may be applied to the cancer patient.

[00035] It will be appreciated that the disclosed methods and/or block diagrams may be implemented as computer-executable instructions, in some embodiments. Thus, in one example, a computer-readable storage device e.g., a non-transitory computer-readable medium) may store computer executable instructions that if executed by a machine (e.g., computer, processor) cause the machine to perform the disclosed methods and/or block diagrams. While executable instructions associated with the disclosed methods and/or block diagrams are described as being stored on a computer-readable storage device, it is to be appreciated that executable instructions associated with other example disclosed methods and/or block diagrams described or claimed herein may also be stored on a computer-readable storage device.

[00036] Fig. 3 illustrates some embodiments of a biochemical recurrence (BCR) assessment system 300 configured to make a medical prognosis relating to BCR in a cancer patient.

[00037] The BCR assessment system 300 comprises a memory 101 configured to store digitized pathology imaging data 102. The digitized pathology imaging data 102 may include imaging data from a cancer patient 302 that has undergone a radical prostatectomy to treat prostate cancer. In some embodiments, the digitized pathology imaging data 102 comprises one or more digitized pathology images 104 (e.g., one or more digitized biopsy slides) obtained from a pathological tissue sample taken from the prostate of the cancer patient 302. In some embodiments, the one or more digitized pathology images 104 may comprise one or more of a whole side image (WSI), patches of a WSI, or the like. In some embodiments, the memory 101 may comprise electronic memory (e.g., solid state memory, SRAM (static randomaccess memory), DRAM (dynamic random-access memory), and/or the like). [00038] In some embodiments, the one or more digitized pathology images 104 may be generated by an image generation stage 303 that is configured to digitize a stained slide generated from a tissue sample 306 taken from the cancer patient 302. In some embodiments, the image generation stage 303 may comprise a tissue resection tool 304 (e.g., a scalpel, a needle, scissors, a biopsy punch, and/or the like) that is used to surgically excise a tissue sample 306 comprising prostate tissue from the cancer patient 302. The tissue sample 306 is provided to a tissue sectioning and staining tool 308, which is configured to slice the tissue sample 306 into thin slices that are placed on one or more transparent slides (e.g., one or more glass slides). The tissue on the one or more transparent slides is then stained to generate one or more tissue slides. The one or more tissue slides are subsequently converted to the one or more digitized pathology images 104 by a slide digitization tool 310 (e.g., comprising a CMOS image sensor, a CCD camera, and/or the like). In some embodiments, the one or more digitized pathology images 104 may comprise a whole slide image of a H&E (Hematoxylin and Eosin) stained slide. [00039] A segmentation tool 106 may be configured to access the one or more digitized pathology images 104. In some embodiments, the segmentation tool 106 comprises a first segmentation stage 108 and a second segmentation stage 112 downstream of the first segmentation stage 108. The first segmentation stage 108 is configured to segment the one or more digitized pathology images 104 to identify one or more tumor regions 110 within the one or more digitized pathology images 104. The second segmentation stage 112 is configured to segment the one or more tumor regions 110 to identify one or more of glandular regions 114, invasive cribriform adenocarcinoma (ICC) 116, and immune cells 118 within the one or more tumor regions 110. In some embodiments, the one or more glandular regions 114 may comprise glandular lumina. In some embodiments, the immune cells may comprise nuclei 314 including tumor infiltrating lymphocyte (TIL) nuclei 316 and/or non-TIL nuclei 318. In some embodiments, the segmentation tool 106 is configured to generate one or more binary masks that identify the glandular lumina 312, the ICC 116, and/or the nuclei 314. In some such embodiments, the one or more binary masks comprise images having a value of “1 ” in image units (e.g., pixels, voxels, etc.) identified as being within the glandular lumina 312, the ICC 116, and/or the nuclei 314 and having a value of “0” in image units outside of the glandular lumina 312, the ICC 116, and/or the nuclei 314.

[00040] In some embodiments, the first segmentation stage 108 and the second segmentation stage 112 may comprise machine learning models. For example, the first segmentation stage 108 may comprise a first machine learning model that has been trained to perform tumor segmentation. In some embodiments, the first machine learning model may comprise a U-net model. The second segmentation stage 112 may comprise a second machine learning model that has been trained to perform segmentation of one or more of the glandular lumina 312, the ICC 116, and the nuclei 314. In some embodiments, the second machine learning model may comprise a HoVer-Net model. In some embodiments, the first machine learning model and the second machine learning model may be implemented as computer code run on one or more processors (e.g., a central processing unit including one or more transistor devices configured to operate computer code to achieve a result, a microcontroller, a graphics processing unit (GPU), or the like).

[00041] A feature extraction tool 120 is configured to extract a plurality of features 121 from the one or more digitized pathology images 104. In some embodiments, the plurality of features 121 may comprise one or more of immune cell spatial features 122, glandular morphometric features 124, and ICC area features 126. In some embodiments, the immune cell spatial features 122 may comprise one or more of nuclei spatial architecture features 320. In some embodiments, the plurality of features 121 may comprise statistical measures (e.g., mean, median, skew, kurtosis, standard deviation, and/or the like) taken over a plurality of regions (e.g., over a plurality of patches) of a digitized pathology image. In some embodiments, the feature extraction tool 120 may be implemented as computer code run by a processing unit (e.g., a CPU, a microcontroller, a GPU, or the like).

[00042] In some embodiments, the nuclei spatial architecture features 320 comprise features that characterize a spatial arrangement between and/or interactions between the TIL nuclei 316 and the non-TIL nuclei 318. In some embodiments, the nuclei spatial architecture features 320 may be extracted from cell graphs/subgraphs and/or cell clusters formed using the TIL nuclei 316 and the non- TIL nuclei 318. In some embodiments, the nuclei spatial architecture features 320 comprise features related to TIL and non-TIL cluster densities, features related to cluster arrangements and neighborhoods (e.g., local interactions of TIL and non-TIL nuclei clusters), and/or the like. In some embodiments, the glandular morphometric features 124 comprise features that characterize a size and/or shape of a gland within a prostate and/or a lumen of a gland within the prostate. For example, the glandular morphometric features 124 may comprise features related to lumen irregularity and uniformity. In some embodiments, the ICC area features 126 may comprise or be an ICC area within a tumor region, a cribriform area index (e.g., a fraction of a tumor area composed of ICC), and/or the like.

[00043] The plurality of features 121 are provided to a machine learning model 128 that has been trained to generate a medical prognosis 130. The medical prognosis 130 may relate to BCR in the cancer patient 302. In some embodiments, the machine learning model 128 may be configured to perform survival analysis to generate a risk score (e.g., corresponding to a risk of BCR for the cancer patient 302). The machine learning model 128 is further configured to compare the risk score to a risk score threshold to identify the cancer patient 302 as being a low-risk 132 for BCR or as being a high-risk 134 for BCR. In some embodiments, the machine learning model 128 may comprise a Cox proportional hazards model implemented as computer code run on one or more processors (e.g., including a CPU, a microcontroller, a GPU, or the like). The medical prognosis 130 can be utilized by health care professionals to make a more informed decision relating to the treatment of the cancer patient 302, thereby allowing for the cancer patient 302 to have an improved quality of life (e.g., a lower risk of death, to avoid negative side effects of drugs that are not likely to produce positive outcomes, and/or the like).

[00044] It will be appreciated that the use of different features may provide for different accuracies in the medical prognosis generated by the disclosed BCR assessment system. Figs. 4-6 illustrate graphs showing performance metrics of disclosed BCR assessment systems trained using different combinations of features. As can be seen in Figs. 4-6, the BCR assessment systems trained using a combination of cribriform area index, glandular features, and immune spatial features achieved a highest accuracy for medical prognoses of BCR.

[00045] Fig. 4 illustrates a graph 400 showing exemplary Kaplan- Meier (KM) curves associated with a probability of BCR-free survival of a prostate cancer patient as a function of time for a disclosed BCR assessment system that uses a cribriform area index (e.g., a fraction of a tumor area composed of invasive cribriform adenocarcinoma (ICC)) as a sole feature for generation of a medical prognosis.

[00046] In graph 400, the x-axis represents a time in months and y-axis represents an estimated BCR-free survival probability. The KM curves include a high-risk group 402 and a low-risk group 404. A stratification between the high-risk group 402 and the low-risk group 404 is based on a median of risk scores generated by the disclosed BCR assessment system during testing. The graph 400 shows significant differences in the KM curves for the high-risk group 402 and the low-risk group 404, thereby indicating a prognostic value of the cribriform area index in generating a medical prognosis relating to BCR for a prostate cancer patient. In some embodiments, the disclosed BCR assessment system may achieve a hazard ratio of approximately 1 .58 with a p-value of 0.0805. In some embodiments, the hazard ratio may vary between 0.938 and 2.68 over a data set.

[00047] Fig. 5A illustrates a graph 500 showing exemplary KM curves associated with a probability of BCR-free survival of a prostate cancer patient as a function of time for a disclosed BCR assessment system that uses a combination of cribriform area index and glandular morphometric features for generation of a medical prognosis.

[00048] In graph 500, the x-axis represents a time in months and y-axis represents an estimated BCR-free survival probability. The KM curves include a high-risk group 502 and a low-risk group 504. A stratification between the high-risk group 502 and the low-risk group 504 is based on a median of risk scores generated by the disclosed BCR assessment system during testing.

[00049] The graph 500 shows significant differences in the KM curves for the high- risk group 502 and the low-risk group 504, thereby indicating a prognostic value of the combination of cribriform area index and glandular morphometric features in generating a medical prognosis relating to BCR for a prostate cancer patient. In some embodiments, the disclosed BCR assessment system may achieve a hazard ratio of approximately 2.18 with a p-value of 0.0863. In some embodiments, the hazard ratio may vary between 1 .24 and 3.84 over a data set. The hazard ratio associated with the combination of cribriform area index and glandular morphometric features is higher than that of cribriform area index alone (e.g., as shown in Fig. 4), thereby indicating that a combination of cribriform area index and glandular morphometric features improves an accuracy of a medical prognosis generated by the disclosed BCR assessment system.

[00050] Fig. 5B illustrates a graph 506 showing an exemplary receiver operating characteristic (ROC) curve for the disclosed BCR assessment system trained using a combination of cribriform area index and glandular morphometric features.

[00051] The graph 506 shows a true positive rate along a y-axis and a false positive rate along an x-axis. A disclosed BCR assessment system trained using a combination of cribriform area index and glandular features achieves a ROC curve 508. The average area under the ROC curve 508 (AUG) was 0.72. The relatively high AUC of the ROC curve 508 further suggests that the use of a combination of cribriform area index and glandular morphometric features can be used to accurately predict BCR in patients.

[00052] Fig. 6 illustrates a graph 600 showing exemplary KM curves associated with a probability of BCR-free survival of a prostate cancer patient as a function of time for a disclosed BCR assessment system that uses a combination of cribriform area index, glandular morphometric features, and immune spatial features for generation of a medical prognosis.

[00053] In graph 600, the x-axis represents a time in months and y-axis represents an estimated BCR-free survival probability. The KM curves include a high-risk group 602 and a low-risk group 604. A stratification between the high-risk group 602 and the low-risk group 604 is based on a median of risk scores generated by the disclosed BCR assessment system during testing.

[00054] The graph 600 shows significant differences in the KM curves for the high- risk group 602 and the low-risk group 604, thereby indicating a prognostic value of the combination of cribriform area index, glandular morphometric features, and immune spatial features in generating a medical prognosis relating to BCR for a prostate cancer patient. In some embodiments, the disclosed BCR assessment system may achieve a hazard ratio of approximately 2.31 with a p-value of 0.0029 and a confidence interval of 95%. In some embodiments, the hazard ratio may vary between 1.35 and 3.97 over a data set. The hazard ratio associated with the combination of the cribriform area index, the glandular morphometric features, and the immune spatial features is higher than that of cribriform area index alone (e.g., as shown in Fig. 4) and that of the combination of cribriform area index and glandular morphometric features (e.g., as shown in Fig. 5A), thereby indicating that a combination of cribriform area index, glandular morphometric features, and immune spatial features improves an accuracy of a prognosis generated by the disclosed BCR assessment system.

[00055] Therefore, the graphs illustrated in Figs. 4-6 show that the disclosed BCR assessment system is able to generate an accurate medical prediction of BCR using different combinations of cribriform area index, glandular morphometric features, and immune spatial features. The ability of the disclosed BCR assessment system to generate an accurate medical prediction of BCR improves a computer’s ability to analyze medical images in a manner that accurately identifies cancer patients that are likely to experience BCR. The improved ability to accurately identify cancer patients that are likely to experience BCR can improve treatment of the cancer patients.

[00056] Fig. 7 illustrates some additional embodiments of a block diagram of a BCR assessment system 700 configured to make a medical prognosis relating to BCR for a cancer patient.

[00057] The BCR assessment system 700 comprises a memory 101 configured to store digitized pathology imaging data 102 including a plurality of digitized pathology images 104 from cancer patients. In some embodiments, the cancer patients are prostate cancer patients that have undergone radical prostatectomies. In various embodiments, the plurality of digitized pathology images 104 may be obtained by an image generation stage 303 and/or from an on-line database 702 and/or archive containing digitized pathology images from cancer patients generated at different sites (e.g., different hospitals, research laboratories, and/or the like).

[00058] In some embodiments, prior to including digitized pathology images within the digitized pathology imaging data 102 a pre-processing stage 704 may be configured to operate upon the one or more digitized pathology images 104. The pre-processing stage 704 may be configured to normalize image characteristics (e.g., color, brightness, contrast, etc.) so as to mitigate batch effects (e.g., differences between images obtained from different sites). The pre-processing stage 704 may also be configured to break a WSI into a plurality of non-overlapping patches that cover the WSI. In some embodiments, the plurality of patches may be subsequently stored as part of the digitized pathology imaging data 102. In some embodiments, the pre-processing stage 704 may discard patches that have poor image quality and/or an insufficient amount of tissue (e.g., less than 50% prostate tissue).

[00059] The plurality of digitized pathology images 104 may include a training set 104t and a validation set 104v. The training set 104t comprises digitized pathology images from a first plurality of patients. The validation set 104v comprises digitized pathology images from a second plurality of patients. The training set 104t may be used to train a downstream machine learning model. For example, the training set 104t may be used to train a downstream segmentation tool 106 to perform segmentations that identify tumor regions, glandular lumina, ICC, and nuclei (e.g., TIL nuclei and non-TIL nuclei) and/or a downstream machine learning model 128 to generate a medical prognosis 130 relating to BCR. The validation set 104v may be used to subsequently validate the results of the segmentation tool 106 and/or the machine learning model 128.

[00060] In some embodiments, the machine learning model 128 may be configured to generate a first plurality of risk scores using the training set 104t and/or the validation set 104v. The machine learning model 128 may be further configured to identify a median risk score of the first plurality of risk scores and to subsequently set the median risk score to be a risk score threshold 706 that differentiates between patients categorized as low-risk and high-risk of BCR. The machine learning model 128 may be configured to utilize the risk score threshold 706 to determine if risk scores corresponding to subsequent cancer patients are categorized as being high- risk or low-risk of BCR. For example, cancer patients having a risk score that is less than the risk score threshold 706 may be categorized as low-risk and cancer patients having a risk score that is greater than the risk score threshold 706 may be categorized as high-risk.

[00061] In some embodiments, the machine learning model 128 may include a feature selector 708 configured to select a set of most prognostic features to generate the medical prognosis 130. For example, the features extraction tool 120 may extract a first number of features and then the feature selector 708 may select a smaller second number of the features that are most prognostic of BCR (e.g., that have a most significant impact in determining a risk of BCR). In some embodiments, the first number of features may include glandular lumina features, ICC area features, and nuclei spatial arrangement features and the second number of features may also include glandular lumina features, ICC area features, and nuclei spatial arrangement features.

[00062] In some embodiments, the feature selector 708 may be configured to select the most prognostic features to be less than 50 features, less than 25 features, and/or the like. In some embodiments, the feature selector 708 may be configured to select a most prognostic 16 features. In some such embodiments, 12 of the most prognostic features may be related to the density and local interactions of TIL and non-TIL nuclei clusters, 3 of the most prognostic features may be related to glandular lumen irregularity and uniformity, and one of the most prognostic features may be related to the area of ICC within a tumor.

[00063] In some embodiments, the most prognostic features may be used to train and validate the machine learning model 128. In some embodiments, the machine learning model 128 may comprise a Cox regression model (e.g., a Cox proportional hazards model). In some embodiments, the Cox regression model may comprise a LASSO (least absolute shrinkage and selection operator) algorithm (e.g., a LASSO Cox regression model) that is configured to operate as the feature selector 708. [00064] Fig. 8 illustrates some embodiments of a block diagram of an apparatus 800 configured to make a medical prognosis relating to a cancer patient.

[00065] The apparatus 800 comprises an assessment apparatus 802. The assessment apparatus 802 is coupled to an image generation stage 303, which is configured to generate one or more digitized pathology images of a prostate tissue sample collected from a cancer patient 302.

[00066] The assessment apparatus 802 comprises a processor 806 and a memory 804. The processor 806 can, in various embodiments, comprise circuitry such as, but not limited to, one or more single-core or multi-core processors. The processor 806 can include any combination of general-purpose processors and dedicated processors (e.g., graphics processors, application processors, etc.). The processor(s) 806 can be coupled with and/or can comprise memory (e.g., memory 804) or storage and can be configured to execute instructions stored in the memory 804 or storage to enable various apparatus, applications, or operating systems to perform operations and/or methods discussed herein. [00067] The memory 804 can be configured to store digitized pathology imaging data 102 including one or more digitized pathology images 104. The one or more digitized pathology images 104 may comprise digitized biopsy images having a plurality of pixels, each pixel having an associated intensity. In some additional embodiments, the one or more digitized pathology images 104 may be stored in the memory 804 as one or more training sets for training a classifier and/or one or more test sets (e.g., validation sets).

[00068] The assessment apparatus 802 also comprises an input/output (I/O) interface 808 (e.g., associated with one or more I/O devices), a display 810, and an interface 812 that connects the processor 806, the memory 804, and the I/O interface 808. The I/O interface 812 can be configured to transfer data between the memory 804, the processor 806, and external devices, for example, the image generation stage 303.

[00069] In some embodiments, the assessment apparatus 802 may further comprise one or more circuits 814 that include one or more of a segmentation circuit 815, a feature extraction circuit 823, and a machine learning circuit 824. In some embodiments, the one or more circuits 814 may operate according to computer code to implement algorithms stored in the memory 804.

[00070] In some embodiments, the segmentation circuit 815 is configured to segment the one or more digitized pathology images 104 to generate segmented digitized pathology images 816 that identify one or more of glandular regions 818, ICC 820, and immune cells 822 (e.g., TIL nuclei) within the one or more digitized pathology images 104. In some embodiments, the segmented digitized pathology images 816 may comprise binary masks. The feature extraction circuit 823 is configured to extract a plurality of features 121 from the segmented digitized pathology images 816 using one or more of the glandular regions 818, the ICC 820, and the immune cells 822. In some embodiments, the feature extraction circuit 823 may be configured to generate one or more of immune cell spatial features 122 (e.g., nuclei spatial arrangement features), glandular morphometric features 124, and ICC area features 126. The machine learning circuit 824 is configured to utilize the plurality of features 121 to generate a medical prognosis 130 relating to the cancer patient r02 (e.g., a medical prognosis relating to BCR in the cancer patient 302). In some embodiments, the display 810 is configured to output or display the medical prognosis 130 generated by the assessment apparatus 802.

Example use case:

[00071] Men with prostate cancer (PCa) that experience biochemical recurrence (BCR) post-radical prostatectomy (RP) are at elevated risk for metastasis and prostate cancer (PCa) specific mortality. Identifying risk of BCR early can help identify men who might benefit from adjuvant therapies like radiation or docetaxel. In this study we sought to identify patients with PCa at high risk of 5-year post-RP BCR utilizing an Artificial Intelligence-based computational pathology model which involves extracting features relating to both glandular morphology and spatial architecture of Tumor-Infiltrating Lymphocytes (TIL) from images of RP specimens. [00072] Methods: Two separate cohorts of RP specimens were collected from the University of Pennsylvania: Si for training (170 patients) and S2 for validation (171 patients). None of the patients underwent any pre-operative treatment. Using two machine learning models (U-net and HoVer-Net), we segmented the tumor area on the digitized pathology slides, from which we further isolated glandular lumina, invasive cribriform adenocarcinoma (ICC), and nuclei. We automatically identified TIL and non-TIL nuclei, derived 350 features characterizing their spatial arrangement, extracted 191 morphometric features from gland lumina, and separately quantified the ICC area. Feature selection was conducted through cross- validation using a least absolute shrinkage and selection operator (LASSO). A Cox model trained on 16 selected features stratified Si patients into high- and low-risk groups using a median risk score threshold for BCR (PSA > 0.2 ng/ml post-surgery) within 5 years. The model's performance was subsequently assessed on the blind cohort S2 by stratifying patients according to the median risk score determined in Si. [00073] Results: The model showed significant prognostic performance for BCR on Si (p < 0.005, HR = 2.31 , 95% confidence interval [Cl]: 1 .35-3.97) and S2 (p < 0.01 , HR = 2.39, 95% Cl: 1 .32-4.33). From the selected 16 features, 12 were related to the density and local interactions of TIL and non-TIL nuclei clusters, 3 related to lumen irregularity and uniformity, and one related to the area of ICC within the tumor. [00074] Conclusion: Our prognostic model successfully identified significant features associated with prostate glandular lumina, ICC, and TILs, that were strongly correlated with the risk of BCR within 5 years of RP. Additional independent multisite validation of these findings is warranted as also evaluating the model from baseline biopsies.

[00075] Therefore, the present disclosure relates to a method and apparatus configured to generate a medical prognosis that identifies cancer patients at a high- risk for biochemical recurrence by operating a machine learning model on features extracted from digitized pathology imaging data. The features relate to a glandular morphology, an invasive cribriform adenocarcinoma (ICC) area, and/or a spatial architecture of immune cells (e.g., tumor-infiltrating lymphocytes).

[00076] In some embodiments, the present disclosure relates a method, including accessing digitized pathology imaging data of tissue excised from a cancer patient, the digitized pathology imaging data being segmented to identify one or more of glandular regions, immune cells, and invasive cribriform adenocarcinoma (ICC); extracting a plurality of features from the digitized pathology imaging data, the plurality of features including one or more of immune cell spatial features, glandular morphometric features, and ICC area features; and providing the plurality of features to a machine learning model that is trained to generate a medical prognosis for the cancer patient using the plurality of features. In some embodiments, the digitized pathology imaging data has been segmented to identify tumor infiltrating lymphocyte (TIL) nuclei and non-TIL nuclei; and the immune cell spatial features including features characterizing spatial arrangements of the TIL nuclei and the non-TIL nuclei. In some embodiments, the plurality of features include the immune cell spatial features and the glandular morphometric features. In some embodiments, the plurality of features include the immune cell spatial features, the glandular morphometric features, and the ICC area features. In some embodiments, the glandular regions are glandular lumina. In some embodiments, the digitized pathology imaging data includes one or more digitized pathology images of prostate tissue from the cancer patient having undergone a radical prostatectomy; and the medical prognosis relates to a risk of biochemical recurrence (BCR) for the cancer patient. In some embodiments, the method further includes performing a first segmentation process to identify a tumor region within the digitized pathology imaging data; and performing a second segmentation process to identify one or more of the glandular regions, the immune cells, and the ICC within the tumor region. In some embodiments, the first segmentation process is performed by a first machine learning model and the second segmentation process is performed by a second machine learning model. In some embodiments, the first machine learning model is a U-net and the second machine learning model is a HoVer-Net. In some embodiments, the plurality of features characterize a density and local interactions of TIL nuclei and non-TIL nuclei, glandular lumen irregularity and uniformity, and an area of the ICC within a tumor.

[00077] In other embodiments, the present disclosure relates to a non-transitory computer-readable medium storing computer-executable instructions that, when executed, cause a processor to perform operations, including accessing one or more digitized pathology images of tissue from a cancer patient, the one or more digitized pathology images being segmented to identify glandular lumina, tumor infiltrating lymphocyte (TIL) nuclei, and non-TIL nuclei; extracting a plurality of features from the one or more digitized pathology images using the glandular lumina, the TIL nuclei, and the non-TIL nuclei, the plurality of features including glandular morphometric features and features characterizing a spatial arrangement of the TIL nuclei and the non-TIL nuclei; and providing the plurality of features to a machine learning model that is trained to generate a medical prognosis regarding a risk of biochemical recurrence (BCR) for the cancer patient. In some embodiments, the plurality of features characterize a density and local interactions of the TIL nuclei and the non- TIL nuclei, and glandular lumen irregularity and uniformity. In some embodiments, the one or more digitized pathology images have been segmented to further identify invasive cribriform adenocarcinoma (ICC); and the plurality of features further include ICC area features. In some embodiments, the one or more digitized pathology images include a first plurality of images obtained from a first plurality of cancer patients and a second image obtained from a second cancer patient; and the operations further include generating a first plurality of risk scores using the first plurality of images; identifying a median risk score of the first plurality of risk scores; setting the median risk score to be a risk score threshold that differentiates between patients categorized as low-risk and high-risk of BCR; generating a risk score using the second image; and comparing the risk score to the risk score threshold to determine if the second cancer patient is categorized as being high-risk or low-risk of BCR. In some embodiments, the one or more digitized pathology images include images of prostate tissue obtained from the cancer patient during a radical prostatectomy.

[00078] In yet other embodiments, the present disclosure relates to an apparatus, including a memory configured to store one or more digitized pathology images of tissue from a cancer patient, the one or more digitized pathology images being segmented to identify glandular lumina, tumor infiltrating lymphocyte (TIL) nuclei, and non-TIL nuclei; a feature extraction tool configured to extract a plurality of features from the one or more digitized pathology images using the glandular lumina, the TIL nuclei, and the non-TIL nuclei, the plurality of features including immune cell spatial features, glandular morphometric features, and ICC area features; and a machine learning model configured to utilize the plurality of features to generate a medical prognosis regarding a risk of biochemical recurrence (BCR) for the cancer patient. In some embodiments, the one or more digitized pathology images being segmented to further identify invasive cribriform adenocarcinoma (ICC); and the plurality of features further including an ICC area. In some embodiments, the feature extraction tool is configured to extract a first plurality of features from the one or more digitized pathology images; and a feature selector is configured to select a second plurality of features, from the first plurality of features, to be the plurality of features, the second plurality of features being numerically smaller than the first plurality of features and being features that are most prognostic of BCR within the cancer patient. In some embodiments, the second plurality of features characterize a density and local interactions of the TIL nuclei and the non-TIL nuclei, glandular lumen irregularity and uniformity, and an area of the ICC within a tumor. In some embodiments, the second plurality of features are less than 50 features.

[00079] Examples herein can include subject matter such as an apparatus, including a digital whole slide scanner, a CT system, an MRI system, a personalized medicine system, a CADx system, a processor, a system, circuitry, a method, means for performing acts, steps, or blocks of the method, at least one machine-readable medium including executable instructions that, when performed by a machine (e.g., a processor with memory, an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA), or the like) cause the machine to perform acts of the method or of an apparatus or system, according to embodiments and examples described.

[00080] References to “one embodiment”, “an embodiment”, “one example”, and “an example” indicate that the embodiment(s) or example(s) so described may include a particular feature, structure, characteristic, property, element, or limitation, but that not every embodiment or example necessarily includes that particular feature, structure, characteristic, property, element or limitation. Furthermore, repeated use of the phrase “in one embodiment” does not necessarily refer to the same embodiment, though it may.

[00081] “Computer-readable storage device”, as used herein, refers to a device that stores instructions or data. “Computer-readable storage device” does not refer to propagated signals. A computer-readable storage device may take forms, including, but not limited to, non-volatile media, and volatile media. Non-volatile media may include, for example, optical disks, magnetic disks, tapes, and other media. Volatile media may include, for example, semiconductor memories, dynamic memory, and other media. Common forms of a computer-readable storage device may include, but are not limited to, a floppy disk, a flexible disk, a hard disk, a magnetic tape, other magnetic medium, an application specific integrated circuit (ASIC), a compact disk (CD), other optical medium, a random access memory (RAM), a read only memory (ROM), a memory chip or card, a memory stick, and other media from which a computer, a processor or other electronic device can read. [00082] “Circuit”, as used herein, includes but is not limited to hardware, firmware, software in execution on a machine, or combinations of each to perform a function(s) or an action(s), or to cause a function or action from another logic, method, or system. A circuit may include a software controlled microprocessor, a discrete logic (e.g., ASIC), an analog circuit, a digital circuit, a programmed logic device, a memory device containing instructions, and other physical devices. A circuit may include one or more gates, combinations of gates, or other circuit components. Where multiple logical circuits are described, it may be possible to incorporate the multiple logical circuits into one physical circuit. Similarly, where a single logical circuit is described, it may be possible to distribute that single logical circuit between multiple physical circuits. [00083] To the extent that the term “includes” or “including” is employed in the detailed description or the claims, it is intended to be inclusive in a manner similar to the term “comprising” as that term is interpreted when employed as a transitional word in a claim.

[00084] Throughout this specification and the claims that follow, unless the context requires otherwise, the words 'comprise' and 'include' and variations such as 'comprising' and 'including' will be understood to be terms of inclusion and not exclusion. For example, when such terms are used to refer to a stated integer or group of integers, such terms do not imply the exclusion of any other integer or group of integers.

[00085] To the extent that the term “or” is employed in the detailed description or claims (e.g., A or B) it is intended to mean “A or B or both”. When the applicants intend to indicate “only A or B but not both” then the term “only A or B but not both” will be employed. Thus, use of the term “or” herein is the inclusive, and not the exclusive use. See, Bryan A. Garner, A Dictionary of Modern Legal Usage 624 (2d. Ed. 1995).

[00086] While example systems, methods, and other embodiments have been illustrated by describing examples, and while the examples have been described in considerable detail, it is not the intention of the applicants to restrict or in any way limit the scope of the appended claims to such detail. It is, of course, not possible to describe every conceivable combination of components or methodologies for purposes of describing the systems, methods, and other embodiments described herein. Therefore, the invention is not limited to the specific details, the representative apparatus, and illustrative examples shown and described. Thus, this application is intended to embrace alterations, modifications, and variations that fall within the scope of the appended claims.

Claims

What is claimed is:

1 . A method, comprising: accessing digitized pathology imaging data of tissue excised from a cancer patient, wherein the digitized pathology imaging data has been segmented to identify one or more of glandular regions, immune cells, and invasive cribriform adenocarcinoma (ICC); extracting a plurality of features from the digitized pathology imaging data, wherein the plurality of features include one or more of immune cell spatial features, glandular morphometric features, and ICC area features; and providing the plurality of features to a machine learning model that is trained to generate a medical prognosis for the cancer patient using the plurality of features.

2. The method of claim 1 , wherein the digitized pathology imaging data has been segmented to identify tumor infiltrating lymphocyte (TIL) nuclei and non-TIL nuclei; and wherein the immune cell spatial features include features characterizing spatial arrangements of the TIL nuclei and the non-TIL nuclei.

3. The method of claim 1 , wherein the plurality of features include the immune cell spatial features and the glandular morphometric features.

4. The method of claim 1 , wherein the plurality of features include the immune cell spatial features, the glandular morphometric features, and the ICC area features.

5. The method of claim 1 , wherein the glandular regions are glandular lumina.

6. The method of claim 1 , wherein the digitized pathology imaging data includes one or more digitized pathology images of prostate tissue from the cancer patient having undergone a radical prostatectomy; and wherein the medical prognosis relates to a risk of biochemical recurrence (BCR) for the cancer patient.

7. The method of claim 1 , further comprising: performing a first segmentation process to identify a tumor region within the digitized pathology imaging data; and performing a second segmentation process to identify one or more of the glandular regions, the immune cells, and the ICC within the tumor region.

8. The method of claim 7, wherein the first segmentation process is performed by a first machine learning model and the second segmentation process is performed by a second machine learning model.

9. The method of claim 8, wherein the first machine learning model is a Linet and the second machine learning model is a HoVer-Net.

10. The method of claim 1 , wherein the plurality of features characterize a density and local interactions of TIL nuclei and non-TIL nuclei, glandular lumen irregularity and uniformity, and an area of the ICC within a tumor.

11. A non-transitory computer-readable medium storing computerexecutable instructions that, when executed, cause a processor to perform operations, comprising: accessing one or more digitized pathology images of tissue from a cancer patient, wherein the one or more digitized pathology images have been segmented to identify glandular lumina, tumor infiltrating lymphocyte (TIL) nuclei, and non-TIL nuclei; extracting a plurality of features from the one or more digitized pathology images using the glandular lumina, the TIL nuclei, and the non-TIL nuclei, wherein the plurality of features include glandular morphometric features and features characterizing a spatial arrangement of the TIL nuclei and the non-TIL nuclei; and providing the plurality of features to a machine learning model that is trained to generate a medical prognosis regarding a risk of biochemical recurrence (BCR) for the cancer patient.

12. The non-transitory computer-readable medium of claim 11 , wherein the plurality of features characterize a density and local interactions of the TIL nuclei and the non-TIL nuclei, and glandular lumen irregularity and uniformity.

13. The non-transitory computer-readable medium of claim 11 , wherein the one or more digitized pathology images have been segmented to further identify invasive cribriform adenocarcinoma (ICC); and wherein the plurality of features further include ICC area features.

14. The non-transitory computer-readable medium of claim 11 , wherein the one or more digitized pathology images comprise a first plurality of images obtained from a first plurality of cancer patients and a second image obtained from a second cancer patient; and wherein the operations further include: generating a first plurality of risk scores using the first plurality of images; identifying a median risk score of the first plurality of risk scores; setting the median risk score to be a risk score threshold that differentiates between patients categorized as low-risk and high-risk of BCR; generating a risk score using the second image; and comparing the risk score to the risk score threshold to determine if the second cancer patient is categorized as being high-risk or low-risk of BCR.

15. The non-transitory computer-readable medium of claim 11 , wherein the one or more digitized pathology images comprise images of prostate tissue obtained from the cancer patient during a radical prostatectomy.

16. An apparatus, comprising: a memory configured to store one or more digitized pathology images of tissue from a cancer patient, wherein the one or more digitized pathology images have been segmented to identify glandular lumina, tumor infiltrating lymphocyte (TIL) nuclei, and non-TIL nuclei; a feature extraction tool configured to extract a plurality of features from the one or more digitized pathology images using the glandular lumina, the TIL nuclei, and the non-TIL nuclei, wherein the plurality of features include immune cell spatial features, glandular morphometric features, and ICC area features; and a machine learning model configured to utilize the plurality of features to generate a medical prognosis regarding a risk of biochemical recurrence (BCR) for the cancer patient.

17. The apparatus of claim 16, wherein the one or more digitized pathology images have been segmented to further identify invasive cribriform adenocarcinoma (ICC); and wherein the plurality of features further include an ICC area.

18. The apparatus of claim 17, wherein the feature extraction tool is configured to extract a first plurality of features from the one or more digitized pathology images; and wherein a feature selector is configured to select a second plurality of features, from the first plurality of features, to be the plurality of features, the second plurality of features being numerically smaller than the first plurality of features and being features that are most prognostic of BCR within the cancer patient.

19. The apparatus of claim 18, wherein the second plurality of features characterize a density and local interactions of the TIL nuclei and the non-TIL nuclei, glandular lumen irregularity and uniformity, and an area of the ICC within a tumor.

20. The apparatus of claim 18, wherein the second plurality of features are less than 50 features.