Masked hypergraph learning for weakly supervised histopathology whole slide image classification.

Area of Science:

• Computational pathology
• Machine learning for medical imaging
• Graph neural networks

Background:

• Whole slide images (WSI) in histopathology present complex data challenges.
• Existing graph neural network (GNN) methods for WSI analysis often overlook non-pairwise patch relationships.
• This limitation hinders optimal feature learning and classification performance.

Purpose of the Study:

• To explore and exploit non-pairwise relationships within histopathology WSIs.
• To develop a novel framework for learning slide-level representations.
• To enhance WSI classification performance by integrating non-pairwise patch correlations.

Main Methods:

• Proposed a Masked HyperGraph Learning (MaskHGL) framework for weakly supervised WSI classification.
• Utilized hypergraphs to model non-pairwise patch correlations and employed hypergraph convolution for global message passing.
• Incorporated a masked hypergraph reconstruction module for improved robustness and generalization, alongside a self-attention node aggregator.

Main Results:

• Evaluated MaskHGL on TCGA-LUNG, TCGA-EGFR, and USTC-EGFR datasets.
• Achieved high Area Under the ROC Curve (AUC) values: 0.9752±0.0024 (TCGA-LUNG), 0.7421±0.0380 (TCGA-EGFR), and 0.8745±0.0100 (USTC-EGFR).
• Demonstrated superior performance over state-of-the-art methods, outperforming SlideGraph+ by 2.64% on the USTC-EGFR dataset.

Conclusions:

• MaskHGL significantly improves WSI classification by leveraging non-pairwise relationships.
• The masked hypergraph reconstruction module enhances robustness and classification ability, particularly in data-scarce scenarios.
• The method shows strong potential for cancer subtyping and predicting gene mutations from H&E stained WSIs.