Hepatocellular carcinoma (HCC) is a heterogeneous disease with diverse characteristics and outcomes. Here, we aim to develop a histological classification for HCC by integrating computational imaging features of the tumor and its microenvironment. We first trained a multitask deep-learning neural network for automated single-cell segmentation and classification on hematoxylin- and eosin-stained tissue sections. After confirming the accuracy in a testing set, we applied the model to whole-slide images of 304 tumors in the Cancer Genome Atlas. Given the single-cell map, we calculated 246 quantitative image features to characterize individual nuclei as well as spatial relations between tumor cells and infiltrating lymphocytes. Unsupervised consensus clustering revealed three reproducible histological subtypes, which exhibit distinct nuclear features as well as spatial distribution and relation between tumor cells and lymphocytes. These histological subtypes were associated with somatic genomic alterations (i.e., aneuploidy) and specific molecular pathways, including cell cycle progression and oxidative phosphorylation. Importantly, these histological subtypes complement established molecular classification and demonstrate independent prognostic value beyond conventional clinicopathologic factors. Our study represents a step forward in quantifying the spatial distribution and complex interaction between tumor and immune microenvironment. The clinical relevance of the imaging subtypes for predicting prognosis and therapy response warrants further validation.
Keywords: deep learning; hepatocellular carcinoma; histopathology image; spatial analysis; tumor microenvironment.