Self-supervised learning-based Multi-Scale feature Fusion Network for survival analysis from whole slide images

Area of Science:

• Digital Pathology
• Computational Biology
• Machine Learning in Medicine

Background:

• Prognosis and mortality prediction are crucial for patient treatment evaluation.
• Digital pathology and deep learning enable survival analysis in whole slide images (WSIs).
• Existing methods face challenges with unstable patch sampling, domain gaps in feature extraction, and limited information representation.

Purpose of the Study:

• To develop a novel, stable patch sampling strategy for WSI survival analysis.
• To enhance feature extraction using self-supervised learning.
• To improve survival prediction accuracy by integrating multi-scale information.

Main Methods:

• A new patch sampling strategy based on image information entropy to select representative patches.
• A Multi-Scale feature Fusion Network (MSFN) incorporating a self-supervised feature extractor.
• An attention-based global-local feature fusion mechanism for comprehensive information representation.

Main Results:

• The proposed method demonstrates competitive results on TCGA-GBM and TCGA-LUSC datasets.
• Image information entropy effectively avoids noise from random or blank regions.
• Self-supervised pretraining improves feature extraction efficiency.

Conclusions:

• The novel patch sampling and MSFN approach offers a robust solution for WSI survival analysis.
• This method enhances prediction accuracy by leveraging comprehensive multi-scale information.
• The findings contribute to more reliable prognosis prediction in digital pathology.