TAD-Graph: Enhancing Whole Slide Image Analysis via Task-Aware Subgraph Disentanglement

Area of Science:

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

Background:

• Whole slide images (WSIs) are crucial for cancer diagnosis and prognosis.
• Graph-based methods excel at integrating pathological and contextual features in WSIs.
• High WSI resolution creates large, noisy graphs, leading to shortcut learning and overfitting.

Purpose of the Study:

• To develop a more efficient approach for WSI analysis using graph representations.
• To address challenges of large, noisy graphs in WSI datasets.
• To improve contextual feature extraction for enhanced diagnostic accuracy.

Main Methods:

• Proposed a novel Task-Aware Disentanglement Graph (TAD-Graph) approach.
• Injected stochasticity into WSI graph edge connections.
• Separated WSI graphs into task-relevant and task-irrelevant subgraphs using a graph information bottleneck objective.

Main Results:

• TAD-Graph demonstrated superior performance across three WSI analysis tasks and six benchmark datasets.
• The method effectively disentangled informative subgraphs for enhanced contextual feature extraction.
• Pathological concept-based metrics confirmed improved predictive accuracy and interpretive insights.

Conclusions:

• TAD-Graph offers a more efficient and accurate method for WSI analysis compared to existing approaches.
• The approach enhances contextual feature learning while mitigating issues from large, noisy graphs.
• TAD-Graph provides interpretive insights and aids in potential biomarker identification for cancer diagnosis.