Generating Hypergraph-Based High-Order Representations of Whole-Slide Histopathological Images for Survival...

Area of Science:

• Computational pathology
• Medical image analysis
• Machine learning for survival analysis

Background:

• Accurate patient survival prediction from whole-slide histopathological images (WSIs) is crucial but challenging due to complex data correlations.
• Developing informative, survival-specific global representations from WSIs is a key obstacle.

Purpose of the Study:

• To propose HGSurvNet, a multi-hypergraph learning framework for effective high-order global representation of WSIs for patient survival prediction.
• To address overfitting issues in survival prediction tasks with limited data.

Main Methods:

• Utilized a multi-hypergraph learning framework (HGSurvNet) for multilateral correlation modeling in multiple spaces.
• Introduced hypergraph max-mask convolution to alleviate overfitting.
• Employed Bayesian Concordance Readjust loss for enhanced performance.

Main Results:

• HGSurvNet demonstrated superior performance in patient survival prediction across three carcinoma datasets (LUSC, GBM, NLST).
• Quantitative analysis confirmed consistent outperformance compared to state-of-the-art methods.
• Individual module effectiveness and the framework's interpretability potential were validated.

Conclusions:

• HGSurvNet offers a robust approach for survival prediction using WSIs, effectively modeling complex correlations.
• The framework shows promise for enhancing pathology diagnosis and reporting through interpretable AI.