PATH-31. ARTIFICIAL INTELLIGENCE IN GLIOMA PATHOLOGY IMAGE ANALYSIS FOR RISK PREDICTION

Abstract

Abstract Pathological evaluation of tumor tissue images stained with hematoxylin and eosin (H&E) is pivotal in diagnosis and predictive of outcome, yet only a small fraction of the rich phenotypic information on the slide is currently used for clinical care. In this study, we developed a computational approach based on deep learning to predict overall survival within distinct molecular subtypes of glioma patients and to extract prognostic biomarkers from microscopic images of tissue biopsies. Whole-slide images from 766 unique patients [IDH: 336 IDH-wildtype, 364 IDH-mutant, 1p/19q: 142 1p/19q-codeleted, 620 1p/19q-non-codeleted] were obtained from The Cancer Genome Atlas (TCGA). Sub-images that were free of artifacts and that contained viable tumor with descriptive histologic characteristics were extracted, which were used for training and testing the deep neural-network. Our unified survival deep learning framework (SDL) uses a residual CNN network integrated with a traditional survival model to predict patient risk from digitized whole-slide images. We employed statistical sampling techniques and randomized transformation of images to address challenges in learning from histology images. Univariable and multivariable Cox proportional-hazards regression models were used to evaluate the significance of predicted patient risk with and without controlling for known prognostic factors. The integrated SDL framework showed substantial prognostic power achieving a median c-index of 0.79 [95 % CI 0.77 - 0.81]. In multivariable Cox regression analysis, SDL risk was significantly associated with overall survival (hazard ratio of 1.65, 95% CI 1.49-1.83, p < 0.001) after adjusting for age, grade, IDH status, ATRX status, 1p19q codeletion and CDKN2A/2B status. Only IDH-status and age were also significant in the multivariable model. Preliminary findings highlight the emerging role of AI in precision medicine and suggest the utility for computational analysis of tumor tissue images for objective and accurate prediction of outcome for glioma patients and also for risk stratification for targeted clinical therapy.