Big-Hypergraph Factorization Neural Network for Survival Prediction From Whole Slide Image.

Area of Science:

• Computational pathology
• Digital pathology
• Machine learning in oncology

Background:

• Survival prediction from histopathological whole-slide images (WSIs) is crucial for patient management.
• Gigapixel WSIs contain vast cell-level information, but extracting stromal/tumor microenvironment data is challenging.
• Current graph-based models often rely on limited patch sampling, hindering transductive learning.

Purpose of the Study:

• To develop a novel method for accurate survival prediction from WSIs.
• To overcome the sampling scale limitations inherent in traditional hypergraph models.
• To improve the exploitation of cell-level structural information within gigapixel WSIs.

Main Methods:

• A factorization neural network was proposed to embed correlations into low-dimensional latent spaces, enabling dense sampling.
• Hypergraph convolutional layers utilized compressed embeddings to generate high-order global representations for each WSI.
• A multi-level ranking supervision strategy was implemented for metric-driven learning and to mitigate uncertainty.

Main Results:

• The proposed method demonstrated superior performance across three public carcinoma datasets (LUSC, GBM, NLST).
• Quantitative results showed significant improvements over existing state-of-the-art survival prediction techniques.
• The approach effectively addressed the bottleneck of sampling scale in hypergraph construction.

Conclusions:

• The factorization neural network offers a powerful approach for survival prediction using WSIs.
• The method successfully leverages dense sampling and global representations for enhanced accuracy.
• This work advances computational pathology by enabling more comprehensive analysis of gigapixel histopathological data.