MuRCL: Multi-Instance Reinforcement Contrastive Learning for Whole Slide Image Classification.

Area of Science:

• Digital Pathology
• Computational Biology
• Machine Learning

Background:

• Whole slide image (WSI) analysis commonly uses multi-instance learning (MIL), which faces overfitting challenges due to weak slide-level supervision.
• Existing methods often focus on instance feature extraction, neglecting the potential of inter-instance relationships for improving classification accuracy.

Purpose of the Study:

• To propose a novel framework, Multi-instance Reinforcement Contrastive Learning (MuRCL), to combat overfitting in MIL for WSI classification.
• To exploit latent semantic relationships among instances (patches) within WSIs to enhance model generalizability.
• To advance WSI classification by deeply mining inherent patch relationships.

Main Methods:

• Developed a two-stage framework: self-supervised pre-training followed by fine-tuning with slide-level labels.
• Employed contrastive learning (CL) to construct discriminative feature sets from patch-level bags.
• Integrated a reinforcement learning-based agent to dynamically update feature set selection for slide-level aggregation.

Main Results:

• MuRCL demonstrated superior performance compared to state-of-the-art MIL models on Camelyon16, TCGA-Lung, and TCGA-Kidney datasets.
• Achieved comparable results to existing models on the TCGA-Esca dataset, validating its effectiveness across different WSI classification tasks.

Conclusions:

• The proposed MuRCL framework effectively mitigates overfitting in MIL for WSI analysis.
• Exploiting inter-instance relationships through reinforcement learning and contrastive learning significantly enhances WSI classification generalizability.
• MuRCL offers a promising approach for robust and accurate automated analysis of whole slide images.