Multi-scale relational graph convolutional network for multiple instance learning in histopathology images.

Area of Science:

• Computational pathology
• Graph-based machine learning
• Medical image analysis

Background:

• Graph convolutional neural networks show promise for histopathology images.
• Existing methods often use single magnification or limited multi-magnification graph structures.
• There is a need to effectively integrate information across different magnifications.

Purpose of the Study:

• To introduce the Multi-Scale Relational Graph Convolutional Network (MS-RGCN).
• To leverage multi-magnification information for improved graph convolutional network performance.
• To enhance message passing between different embedding spaces in histopathology image analysis.

Main Methods:

• Developed MS-RGCN as a multiple instance learning method.
• Modeled histopathology image patches and their multi-scale relations as a graph.
• Utilized separate message-passing neural networks for different node and edge types across magnifications.

Main Results:

• MS-RGCN outperformed state-of-the-art methods on prostate cancer histopathology images.
• Achieved superior performance across various datasets and image types (TMAs, WSI regions, WSIs).
• Ablation studies confirmed the effectiveness of key MS-RGCN design features.

Conclusions:

• MS-RGCN effectively integrates multi-magnification information for histopathology image analysis.
• The proposed method demonstrates significant improvements in predicting cancer grade groups.
• MS-RGCN offers a robust framework for advanced computational pathology applications.