TMOD-35. PREDICTION OF OVERALL SURVIVAL, AND MOLECULAR MARKERS IN GLIOMAS VIA ANALYSIS OF DIGITAL PATHOLOGY IMAGES USING DEEP LEARNING

Abstract

Abstract BACKGROUND Microscopic features of brain tumors, such as tumor cell morphology, type/degree of microvascular hyperplasia, mitotic activity, and extent of zonal/geographic necrosis, among many others, are measurable and reflect underlying molecular markers that are predictive of patient prognosis, signifying that quantitative analysis may provide insight into disease mechanics. We developed a computational method to predict overall-survival and molecular markers for brain tumors using deep learning on whole-slide digital images (WSDI). METHODS The WSDI were acquired from TCGA for 663 patients [IDH:333 wildtype, 330 mutants, 1p/19q:201 non-codeleted, 129 codeleted]. A set of 100 region-of-interest each comprising 1024x1024 that contained viable tumor with descriptive histologic characteristics and that were free of artifacts were extracted. A modified version of ResNet with architecture of 50 convolutional layers was used. The network was optimized using stochastic gradient decent optimization method with binary cross-entropy loss. Sigmoid- and linear-activation were, respectively, used as final layer for mutation and survival prediction. Data was divided into training (50%), testing (25%), and validation (25%). RESULTS The model predicted IDH and 1p/19q with an accuracy of 88.92%[sensitivity(se)/ specificity(sp)=87.77/84.35] and 88.23%(se/sp=87.38/88.58), respectively. The accuracy was further improved, when classification was done within homogeneous grades, for IDH [II=90.50%(se/sp=91.52/78.57), III=91.21%(se/sp=91.89/89.47), IV=92.77%(se/sp=77.78/93.51)] and 1p/19q [II=91.51%(se/sp=91.30/91.66), III=92.56(se/sp =93.33,92.04)]. The Pearson correlation coefficient between the predicted scores and overall-survival was 0.79 (p< 0.0001). CONCLUSION Our findings suggest that deep learning techniques can be applied to WSDI for objective, and accurate prediction of mutations and survival. Our approach, when compared with expensive molecular based assays that invariably capture molecular markers from a small part of the tumor and also destroy the tissue, could (i) offer the same service at a reduced price, (ii) enable disease characterization across the entire landscape of the tissue, (iii) be beneficial for tissues inadequate for molecular testing, and (iv) does not need physical shipping of the tissue.