Prediction of Immune and Stromal Cell Population Abundance from Hepatocellular Carcinoma Whole Slide Images Using Weakly Supervised Learning

Abstract

AbstractWeakly supervised multiple instance learning has been proven to have the ability to predict gene signatures related to immunotherapy sensitivity from hepatocellular carcinoma whole slide images. By examination of the most predictive patches highlighted by the algorithm, specific immune cell populations were enriched in three interferon-gamma-related gene signatures. We proposed to use the algorithm to predict the abundance of various immune and stromal cell populations. The correlation and colocalization with each cell population and gene signature were investigated. It was observed that specific immune cells had positive correlations to interferon-gamma and inflammation gene signatures while stromal cells had negative or trivial correlations. Our models of specific gene signatures and cell populations are promising to develop a novel image-based biomarker for immunotherapy.KeywordsImmune cell populationsWhole slide imageMultiple instance learningWeakly supervised learning