Deep learning tumor heterogeneity metric from histopathology images vs next generation sequencing-derived scores for colon cancer prognostication.

Abstract

3537 Background: Tumor heterogeneity is as an important determinant of clinical behavior in many cancer types, with increased heterogeneity thought to confer inferior clinical outcome. Sequencing based assessment of colon cancer has been used to quantify tumor heterogeneity and correlate it with survival, but is sensitive to the mutation calling algorithm utilized. Digitization of histopathology slides allows application of deep learning methods for image analysis. Automated slide annotation and feature extraction provides numeric representations of underlying phenotype from which tumor heterogeneity can be estimated. Here we compare the ability of a deep-learning tumor heterogeneity score (THS) to estimate overall survival in colon cancer patients to estimates produced by bulk sequencing derived mutant-allele tumor heterogeneity (MATH) and copy number variation (CNV) event scores. Methods: Digitized whole slide images (WSIs) from The Cancer Genome Atlas (TCGA) colon adenocarcinoma dataset are processed in a computational pipeline for tissue detection, numerical feature extraction, and cell type classification. WSIs are tiled and local THS calculated from imaging features. Corresponding MATH and CNV derived heterogeneity metrics are calculated from published sequencing data. Kaplan Meier curves are used to compare patient survival stratified by tumor heterogeneity as calculated by the three independent methods. Automated regional annotation identifies intra-tumoral and tumor-stromal boundary tiles and tumor THS is correlated to distance from the boundary. Results: Images from 379 patients (209 right-sided, 152 left-sided tumors) yielded 575,762 tiles. Automated annotation resulted in 272,791 intra-tumoral and 11,340 tumor-stromal boundary tiles. Stratification by imaging derived THS provided significant separation of overall survival curves (p < .01) in contrast to MATH and CNV methods (p > .05 for both). THS significantly separated the curves in right- but not left-sided cancers (p = .03 and .44, respectively). Evaluating THS as a function of distance from the tumor-stromal boundary, right-sided tumors showed significant decrease in THS with increasing distance from tumor edge. Conclusions: Our novel pipeline produced spatially resolved imaging data informed by underlying tumor phenotype without need for pathologist annotation. The resultant THS correlated with outcome and outperformed sequencing-based prognostication methods. The spatial information identified higher heterogeneity at the tumor edge than interior regions of right-sided, but not left-sided tumors, capturing known but poorly understood differences in colon tumors by location. Deep learning image analysis provided reproducible and cost-effective data with great potential both in clinical biomarker discovery and as a research tool.