Computational features of tumor-infiltrating lymphocyte (TIL) architecture

1. A method, comprising:

accessing a digitized hematoxylin and eosin stained slide (H&E slide) image of an African American (AA) patient, wherein the digitized H&E slide image of the AA patient demonstrates one or more indicators of endometrial cancer (EC), and wherein the digitized H&E slide of the AA patient demonstrates tissue from a uterus of the AA patient and at least a portion of a gynecologic tumor;

defining a tumor region in the digitized H&E slide image, wherein the tumor region comprises at least a part of the portion of the gynecologic tumor, and wherein the tumor region comprises a plurality of individual cells;

defining a boundary for each of the plurality of individual cells;

classifying the plurality of individual cells into cell types, wherein the cell types comprise tumor-infiltrating lymphocytes (TILs), non-lymphocyte cells, and cancer cells;

classifying the TILs as stromal TILs or epithelial TILs;

generating a cluster of stromal TILs, wherein the cluster of stromal TILs comprises a subset of stromal TILs that are related to one another based on proximity;

extracting a first plurality of architectural features from the digitized H&E slide image of the AA patient, wherein each of the first plurality of architectural features are at least partially based on the cluster of stromal TILs;

generating a risk score for the AA patient based on the first plurality of architectural features, wherein the risk score is prognostic of overall survival (OS) of the AA patient;

generating a risk group classification for the AA patient, wherein generating the risk group classification comprises classifying the AA patient into either a high risk group or a low risk group based on the risk score, wherein the high risk group indicates a probability that the AA patient will die within a date range is greater than a threshold probability, and wherein the low risk group indicates the probability that the AA patient will die within the date range is less than or equal to the threshold probability; and

displaying the risk group classification of the AA patient.

2. The method of claim 1 , further comprising:

providing the risk score to a machine learning classifier that is trained to predict a response of the AA patient to a treatment plan for the EC;

receiving, from the machine learning classifier, a classification of the AA patient into either a responder group (RG) or a non-responder group (NRG), where the NRG indicates the AA patient will not respond to the treatment plan and the RG indicates that the AA patient will respond to the treatment plan; and

displaying the classification of the AA patient as either in the NRG or in the RG.

3. The method of claim 2 , wherein the treatment plan comprises at least one of chemotherapy and radiation.

4. The method of claim 1 , further comprising:

extracting a second plurality of architectural features from the digitized H&E slide image of the AA patient, wherein each of the second plurality of architectural features are at least partially based on the cluster of stromal TILs; and

selecting a subset of architectural features of the second plurality of architectural features, wherein the subset of architectural features of the second plurality of architectural features are more relevant to predicting OS of AA patients with EC than the other architectural features of the second plurality of architectural features for a predefined feature selection process, and wherein the subset of architectural features defines the first plurality of architectural features.

5. The method of claim 4 , wherein selecting the subset of architectural features of the second plurality of architectural features comprises:

performing a least absolute shrinkage and selection operator (LASSO) technique on the second plurality of architectural features.

6. The method of claim 4 , wherein the first plurality of architectural features are based only on the cluster of stromal TILs.

7. The method of claim 1 , wherein generating the risk score for the AA patient based on the first plurality of architectural features comprises:

assigning a value to each of the architectural features of the first plurality of architectural features;

assigning a weighting coefficient to each of the values; and

combining the values and their respective weighting coefficients linearly to generate the risk score.

8. The method of claim 1 , wherein the cluster of stromal TILs is generated via a graph theory technique.

9. The method of claim 1 , wherein classifying the AA patient into either the high risk group or the low risk group comprises comparing the risk score of the AA patient to a threshold value.

10. The method of claim 9 , wherein classifying the AA patient into either the high risk group or the low risk group further comprises:

classifying the AA patient into the high risk group if the risk score for the AA patient is greater than the threshold value; and

classifying the AA patient into the low risk group if the risk score for the AA patient is less than or equal to the threshold value.

11. A non-transitory computer-readable storage device storing computer-executable instructions that when executed cause a processor to perform operations, the operations comprising:

accessing a digitized hematoxylin and eosin stained slide (H&E slide) image of an (AA) African American patient, wherein the digitized H&E slide image of the AA patient demonstrates one or more indicators of endometrial cancer (EC), and wherein the digitized H&E slide image of the AA patient demonstrates tissue from a uterus of the AA patient and at least a portion of a gynecologic tumor;

defining a tumor region in the digitized H&E slide image of the AA patient, wherein the tumor region comprises at least a part of the portion of the gynecologic tumor, and wherein the tumor region comprises a plurality of individual cells;

classifying the plurality of individual cells into cell types, wherein the cell types comprise tumor-infiltrating lymphocytes (TILs) and non-lymphocyte cells;

classifying the TILs as intratumoral TILs or stromal TILs;

generating one or more clusters of stromal TILs, wherein each of the one or more clusters of stromal TILs comprises a subset of stromal TILs that are related to one another based on proximity;

extracting a plurality of architectural features from the digitized H&E slide image of the AA patient, wherein the plurality of architectural features are at least partially based on the one or more clusters of stromal TILs;

generating a risk score for the AA patient based on the plurality of architectural features, wherein the risk score is prognostic of overall survival (OS) of the AA patient;

providing the risk score to a machine learning classifier that is trained to predict whether the EC of the AA patient is either an aggressive subtype of EC or a non-aggressive subtype of EC;

receiving, from the machine learning classifier, a classification of the EC of the AA patient as either the aggressive subtype of EC or the non-aggressive subtype of EC; and

displaying the classification of the EC of the AA patient.

12. The non-transitory computer-readable storage device of claim 11 , wherein:

each of the plurality of architectural features corresponds to a different architectural feature of the one or more clusters of stromal TILs.

13. The non-transitory computer-readable storage device of claim 12 , wherein generating the risk score comprises:

assigning a plurality of values to the plurality of architectural features, respectively, wherein each of the values of the plurality of values corresponds to a number of times a corresponding architectural feature of the plurality of architectural features is present in the digitized H&E slide image of the AA patient; and

combining, linearly, the plurality of values with a plurality of weighting coefficients, wherein the plurality of weighting coefficients are attached to the plurality of values, respectively.

14. The non-transitory computer-readable storage device of claim 13 , wherein the plurality of architectural features comprises at least four architectural features.

15. The non-transitory computer-readable storage device of claim 14 , wherein the plurality of architectural features are based only on the one or more clusters of stromal TILs.

16. The non-transitory computer-readable storage device of claim 15 , wherein the plurality of architectural features consists of four architectural features.

17. The non-transitory computer-readable storage device of claim 11 , wherein generating the one or more clusters of stromal TILs comprises:

grouping the stromal TILs into corresponding clusters of the one or more clusters of stromal TILs based on a distance in which the stromal TILs are spaced from one another, wherein each stromal TIL of a given cluster of the one or more clusters of stromal TILs is spaced from a neighboring stromal TIL of the given cluster by less than a threshold distance.

18. A non-transitory computer-readable storage device storing computer-executable instructions that when executed cause a processor to perform operations, the operations comprising:

accessing a training dataset of digitized hematoxylin and eosin stained slide (H&E slide) images, wherein the training dataset of digitized H&E slide images comprises a plurality of digitized H&E training slide images of AA patients, wherein each of the digitized H&E training slide images demonstrates tissue from a uterus of a corresponding AA patient and a portion of a gynecologic tumor of the corresponding AA patient;

defining an AA tumor region for each of the digitized H&E training slide images, wherein each of the AA tumor regions comprises a corresponding plurality of AA cells;

defining a boundary for each of the AA cells of each of the corresponding pluralities of AA cells;

classifying each of the AA cells as one cell type of a plurality of cell types, wherein the plurality of cell types comprises tumor-infiltrating lymphocytes (TILs), non-lymphocyte cells, and cancer cells;

for each corresponding plurality of AA cells, generating one or more clusters of AA stromal TILs and one or more clusters of AA epithelial TILs;

for each of the corresponding plurality of AA cells, extracting a first plurality of architectural features from the one or more clusters of AA stromal TILs;

for each of the corresponding plurality of AA cells, extracting a second plurality of architectural features from the one or more clusters of AA epithelial TILs;

refining the first plurality of architectural features and the second plurality of architectural features to a third plurality of architectural features, wherein the architectural features of the third plurality of architectural features are more relevant to predicting overall survival (OS) of the AA patients than the other architectural features of the first plurality of architectural features and the second plurality of architectural features;

generating risk scores for the AA patients, respectively, wherein each of the risk scores for the AA patients is generated based on the third plurality of architectural features of a corresponding digitized H&E slide image of an AA patient of the plurality of digitized H&E training slide images; and

training a machine learning classifier based on the risk scores for the AA patients, wherein the machine learning classifier is trained to predict a difference between an aggressive subtype of endometrial cancer (EC) and a non-aggressive subtype of EC.

19. The non-transitory computer-readable storage device of claim 18 , wherein the third plurality of architectural features comprises architectural features from only the first plurality of architectural features.

20. The non-transitory computer-readable storage device of claim 18 , wherein the operations further comprise:

accessing a digitized H&E slide image of an AA patient of interest (POI), wherein the digitized H&E slide image of the AA POI indicates the AA POI has EC, and wherein the digitized H&E slide image of the AA POI demonstrates tissue from a uterus of the AA POI and at least a part of the portion of a gynecologic tumor of the AA POI;

defining a tumor region in the digitized H&E slide image of the AA POI, wherein the tumor region comprises a plurality of AA POI cells;

defining a boundary for each of the plurality of AA POI cells;

classifying each of the plurality of AA POI cells as one cell type of the plurality of cell types;

generating one or more clusters of AA POI stromal TILs;

extracting the third plurality of architectural features from the one or more clusters of AA POI stromal TILs;

generating a risk score for the AA POI based on the third plurality of architectural features that were extracted from the one or more clusters of AA POI stromal TILs;

providing the risk score for the AA POI to the machine learning classifier;

receiving, from the machine learning classifier, a classification of the EC of the AA POI into either the aggressive subtype of EC or the non-aggressive subtype of EC; and

displaying the classification of the EC of the AA POI.