Incorporating intratumoral heterogeneity into computational pathology deep learning framework

Predicting cancer survival from gigapixel whole slide images (WSIs) is a critical challenge in computational pathology that requires modeling complex features of the tumor microenvironment. Although there have been advances in using AI models for cancer prognosis, these models do not explicitly capture intratumoral heterogeneity. Using the first two statistical moments (mean and variance) of the features extracted from the WSIs, we develop a novel AI architecture that incorporates intratumoral heterogeneity into its predictions. An empirical study with 4,479 gigapixel WSIs from the Cancer Genome Atlas shows that our proposed model improves survival prediction performance for the five representative cancer types (BRCA, BLCA, GBMLGG, UCEC, and COADREAD). We also present interpretability tools to further probe the biological signals captured by our model. We show that our model can identify/delineate interpretable features, such as cellularity or cell types, within each WSI.