MG-Trans: Multi-Scale Graph Transformer With Information Bottleneck for Whole Slide Image Classification.

Area of Science:

• Digital pathology
• Computational pathology
• Machine learning in healthcare

Background:

• Multiple instance learning (MIL) is standard for whole slide image (WSI) analysis.
• Current MIL methods generate redundant patch data and inadequately model spatial relationships, limiting fine-grained feature discrimination.

Purpose of the Study:

• To address limitations in current MIL methods for WSI classification.
• To propose an effective model for whole slide image classification that enhances feature representation and spatial relationship modeling.

Main Methods:

• Developed Multi-scale Graph Transformer (MG-Trans) with three modules: Patch Anchoring Module (PAM), Structure Information Learning Module (SILM), and Multi-scale Information Bottleneck Module (MIBM).
• PAM samples informative patches using class attention maps from Vision Transformer.
• SILM incorporates local tissue structure for spatial relation modeling.
• MIBM fuses multi-scale features using information bottleneck principle for compact representation.
• Introduced a semantic consistency loss for stable training.

Main Results:

• MG-Trans demonstrated superior performance on WSI classification tasks.
• The model effectively identified informative patches and modeled spatial relationships.
• Multi-scale feature fusion resulted in robust and compact bag-level representations.

Conclusions:

• MG-Trans offers a significant advancement over existing MIL methods for WSI classification.
• The proposed architecture enhances the discriminative ability for fine-grained features in digital pathology.
• MG-Trans shows strong potential for improving diagnostic accuracy in subtyping and gene mutation detection.