A Hierarchical Multi-Scale Graph–Hypergraph Learning Framework for Intelligent Pathology Image Classification

doi:10.1109/tce.2026.3722706

A Hierarchical Multi-Scale Graph–Hypergraph Learning Framework for Intelligent Pathology Image Classification

2026-08-11

0

PRE

AI

DOI：10.1109/tce.2026.3722706

Original

Original request for help

Save

Abstract

En 中文

Whole Slide Image (WSI) classification is a key enabler for scalable digital pathology services in connected healthcare, where consumer-oriented workflows demand both trustworthy predictions and timely time-to-result. However, gigapixel WSIs, weak supervision, and complex multi-scale dependencies make efficient and interpretable modeling challenging. We propose MGH-MIL, a hierarchical multi-scale multiple instance learning framework that integrates graph neural networks and heterogeneous hypergraphs to jointly capture intra- and inter-scale dependencies. Multi-magnification patch features are extracted by a self-supervised encoder, and a quadtree hierarchy defines cross-scale parent–child relations. Within each scale, attention-based intra-scale homogeneous graphs model local spatial correlations, while a cross-scale heterogeneous hypergraph encodes hierarchical and contextual interactions across magnifications. A knowledge distillation objective aligns representations across scales to improve consistency and robustness. Experiments on Camelyon16 and TCGA-Lung show that MGH-MIL achieves superior accuracy and AUC compared with state-of-the-art MIL methods, with favorable inference time. These results suggest that MGH-MIL provides an efficient and interpretable algorithmic component for large-scale, service-oriented digital pathology and consumer-oriented connected healthcare applications.

Keywords:

Whole slide image

graph neural networks

multiscale analysis

attention mechanism

classification

A Hierarchical Multi-Scale Graph–Hypergraph Learning Framework for Intelligent Pathology Image Classification

Abstract

Journal

Organization

Cited Papers