H^2-MIL: Exploring Hierarchical Representation with Heterogeneous Multiple Instance Learning for Whole Slide Image Analysis

Current representation learning methods for whole slide image (WSI) with pyramidal resolutions are inherently homogeneous and flat, which cannot fully exploit the multiscale heterogeneous diagnostic information of different structures comprehensive analysis. This paper presents a novel graph neural network-based multiple instance framework (i.e., H^2-MIL) to learn hierarchical from WSI A “resolution” attribute is constructed explicitly model feature spatial-scaling relationship multi-resolution patches. We then design resolution-aware attention convolution (RAConv) block compact yet discriminative graph, tackles heterogeneity node neighbors yields more reliable message passing. More importantly, explore task-related structured pyramid, we elaborately iterative pooling (IHPool) module progressively aggregate based on scaling relationships nodes. evaluated our method two public datasets TCGA project, i.e., esophageal cancer kidney cancer. Experimental results show that clearly outperforms state-of-the-art both tumor typing staging tasks.