NIMG-60. PREDICTING OVERALL SURVIVAL IN GLIOBLASTOMA USING HISTOPATHOLOGY VIA AN END-TO-END DEEP LEARNING PIPELINE: A LARGE MULTI-COHORT STUDY

Abstract

Abstract PURPOSE Glioblastoma is an aggressive and universally fatal tumor. Morphological information as captured from cellular regions on surgically resected histopathology slides has the ability to reveal the inherent heterogeneity in Glioblastoma and thus has prognostic implications. In this work, we hypothesized that capturing morphological attributes from high cellularity regions on Hematoxylin and Eosin (H&E)-stained digitized tissue slides using an end-to-end deep-learning pipeline will enable risk-stratification of GBM tumors based on overall survival. METHODS A large multi-cohort study consisting of N=514 H&E-stained digitized tissue slides along with overall-survival data (OS) was obtained from the Ivy Glioblastoma atlas project (Ivy-GAP (N=41)), TCGA (N=379), and CPTAC (N=94). Our deep-learning pipeline consisted of two stages. First stage involved segmenting cellular tumor (CT) from necrotic-regions and background using Resnet-18 model, while the second stage involved predicting OS, using only the segmented CT regions identified in the first stage. For the segmentation stage, we leveraged the Ivy-GAP cohort, where CT annotations confirmed by expert neuropathologists were available, to serve as the training set. Using this training model, the CT regions on the remaining cohort (TCGA, CPTAC) (i.e. test set) were identified. For the survival-prediction stage, the last layer of ResNet18 model was replaced with a cox layer (ResNet-Cox), and further fine-tuned using OS and censor information. Independent validation of ResNet-Cox was performed on two hold-out sites from TCGA and one from CPTAC. RESULTS Our segmentation model achieved an accuracy of 0.89 in reliably identifying CT regions on the validation data. The segmented CT regions on the test cohort were further confirmed by two experts. Our ResNet-Cox model achieved a concordance-index of 0.73 on MD Anderson Cancer Center (N=60), 0.71 on Henry Ford Hospital (N=96), and 0.68 on CPTAC data (N=41). CONCLUSION Deep-learning features captured from cellular tumor of H&E-stained histopathology images may predict survival in Glioblastoma.