Diagnose Like A Pathologist: Weakly-Supervised Pathologist-Tree Network for Slide-Level Immunohistochemical Scoring

The immunohistochemistry (IHC) test of biopsy tissue is crucial to develop targeted treatment and evaluate prognosis for cancer patients. IHC staining slide usually digitized into the whole-slide image (WSI) with gigapixels quantitative analysis. To perform a whole prediction (e.g., scoring, survival prediction, grading) from this kind high-dimensional image, algorithms are often developed based on multi-instance learning (MIL) framework. However, multi-scale information WSI associations among instances not well explored in existing MIL studies. Inspired by fact that pathologists jointly analyze visual fields at multiple powers objective diagnostic predictions, we propose Pathologist-Tree Network (PTree-Net) sparsely model efficiently manner. Specifically, Focal-Aware Module (FAM) can approximately estimate diagnosis-related regions an extractor trained using thumbnail WSI. With initial regions, hierarchically patches tree structure, where both global local be captured. explore structure end-to-end network, patch Relevance-enhanced Graph Convolutional (RGCN) explicitly correlations adjacent parent-child nodes, accompanied relevance exploit implicit contextual distant nodes. In addition, tree-based self-supervision devised improve representation suppress irrelevant adaptively. Extensive experiments performed large-scale HER2 dataset. ablation study confirms effectiveness our design, approach outperforms state-of-the-art large margin.