Learning Heterogeneous Embedding with Prototype-Aware Graph Attention for Whole Slide Image Classification

Whole Slide Images (WSIs) are the digital version of pathology slides, central to computational pathology. The WSI pyramid makes it can offer a wide range of diagnostic information, from global tissue structures to detailed cellular features. However, current multi-instance and graph representation learning methods struggle to create a unified framework that effectively captures both local spatial awareness and global WSI representation, limiting their performance in critical tasks such as tumor staging. To this end, we propose a Prototypeaware Heterogeneous Graph ATtention (PHGAT) network that enables each region within a WSI to perceive the representations of its diverse heterogeneous neighbors. This, in turn, guides the learning of WSI-level heterogeneous embedding through multilevel prototypes. Specifically, we introduce three node relations (i.e., local, non-local, and hierarchical) into WSI heterogeneous graph construction and design a novel Heterogeneous Calibration Graph ATtention (HC-GAT) module to propagate the various heterogeneous neighbor node representations within graphs. Then, a Level-aware Prototype Attention module is proposed to obtain prototypes from different levels by aggregating the node representations via a set of trainable query embeddings. Lastly, based on these learned prototypes, a prototype-aware hierarchical pooling module is designed to generate the final heterogeneous embedding of each WSI. Extensive experiments on six diverse datasets across three cancer types and two specific diagnostic tasks show that the proposed framework significantly outperforms the state-of-the-art tumor staging methods and performs comparably in cancer subtyping.