Leveraging commonality across multiple tissue slices for enhanced whole slide image classification using graph...

Area of Science:

• Computational pathology
• Digital pathology
• Machine learning in healthcare

Background:

• Accurate histopathological whole slide image (WSI) classification is crucial for cancer diagnosis and treatment planning.
• Conventional WSI analysis often overlooks shared patterns present across different tissue slices from the same biopsy.
• Deep learning models have shown potential but typically analyze slices independently.

Purpose of the Study:

• To develop a novel WSI classification technique that leverages inter-slice commonality to enhance diagnostic accuracy.
• To improve upon existing deep learning and multiple instance learning approaches for WSI analysis.

Main Methods:

• Constructing graph representations for individual tissue slices.
• Extracting relevant features and connecting graphs based on spatial relationships and feature similarity.
• Utilizing graph convolutional networks for WSI classification on the integrated graph structure.

Main Results:

• The proposed method significantly improves graph-based WSI classification by incorporating inter-slice commonality.
• Achieved higher accuracy (stomach: 91.5%, colorectal: 91.2%) compared to existing methods.
• Demonstrated superior AUROC (stomach: 98.8%, colorectal: 98.2%) compared to multiple instance learning approaches.

Conclusions:

• The novel approach provides a more accurate and efficient method for WSI classification by effectively leveraging information across slices.
• This technique holds significant promise for improving clinical applications in cancer diagnosis.
• Source code is publicly available for further research and development.