Context-Aware Representation Learning for Hepatocellular Carcinoma Prognosis in Whole Slide Images
Luyu Tang, Kun Ru, Wenjian Qin · 2024
Cancer prognosis analysis is a pivotal aspect of clinical practice, which requires a comprehensive consideration of contextual features across different tissue regions to predict patient outcomes. Although existing methods based on multi-instance learning have made substantial advancements across various pathological tasks. However, they often ignore the intricate interactions within pathological images among distinct tissue regions. Moreover, they demonstrate a limited capacity for contextual perception, resulting in suboptimal performance in prognostic assessments. In this study, we proposed a novel contextual information learning method that automatically extracted diverse tissue regions within whole slide images (WSI) from hepatocellular carcinoma patients. The method enabled the extraction of different organizational areas to construct comprehensive organizational structure graphs, which integrated multi-layer perceptron (MLP) mixer into Graph Convolutional Network (GCN-MLP Mixer). Our proposed model was adept at learning the distribution of regions and capturing contextual representations within the tissue region graphs, which can ultimately accurately predict survival outcomes for hepatocellular carcinoma patients. We validated our model in the Cancer Genome Atlas (TCGA) dataset, demonstrating that our method outperformed all prior approaches by 2.6-13.6% in Concordance Index.