Spatially-Constrained and -Unconstrained Bi-Graph Interaction Network for Multi-Organ Pathology Image Classification.

Area of Science:

• Computational pathology
• Medical image analysis
• Graph neural networks

Background:

• Graphs are valuable for pathology image analysis, revealing diverse features.
• Existing graph structures have not been fully integrated for pathology image analysis.
• The interaction between different graph structures remains understudied.

Purpose of the Study:

• To propose SCUBa-Net, a parallel bi-graph neural network for pathology image analysis.
• To leverage both graph convolutional networks and Transformers for enhanced feature extraction.
• To investigate novel inter-graph and intra-graph interaction mechanisms.

Main Methods:

• SCUBa-Net processes pathology images using two distinct graphs: spatially-constrained and spatially-unconstrained.
• Introduced two inter-graph interaction blocks for node-to-node learning within each graph.
• Implemented an intra-graph interaction block for global and local graph-to-graph learning via virtual nodes.

Main Results:

• SCUBa-Net demonstrated effectiveness across four multi-organ cancer datasets (colorectal, prostate, gastric, bladder).
• The proposed interaction blocks significantly improved graph learning efficiency and feature representation.
• Comparative analysis showed superior performance against state-of-the-art CNNs, Transformers, and GNNs.

Conclusions:

• SCUBa-Net offers a powerful new framework for computational pathology image analysis.
• The integration of dual graph structures and interaction blocks advances the field.
• This method holds promise for improving diagnostic accuracy in various cancer types.