Multi-Scale Dynamic Sparse Token Multi-Instance Learning for Pathology Image Classification.

Area of Science:

• Digital Pathology
• Computational Oncology
• Machine Learning in Medicine

Background:

• Whole Slide Images (WSIs) in breast cancer pathology present challenges due to limited informative tumor regions, making subtle lesion identification difficult for pathologists.
• The information gap between diagnostic needs (tumor area < 10%) and WSI data volume necessitates advanced computational approaches.

Purpose of the Study:

• To develop an efficient computational framework for analyzing challenging breast cancer pathology images.
• To address the labor-intensive nature of lesion identification in WSIs.

Main Methods:

• A dynamic sparse token-based multi-instance learning framework with a dynamic sparse layer in the transformer architecture.
• A weakly supervised cross-scale contrastive learning framework leveraging multi-scale pathology image features for bag-level representation.

Main Results:

• The proposed framework demonstrated superior performance across six evaluation metrics compared to state-of-the-art methods.
• Experiments on four cancer datasets validated the model's effectiveness and transferability in both single-scale and multi-scale analyses.

Conclusions:

• The dynamic sparse token and cross-scale contrastive learning framework effectively addresses challenges in breast cancer pathology image analysis.
• The model offers improved accuracy and efficiency for identifying subtle lesions, aiding clinical diagnosis.