Dual-stream Multiple Instance Learning Network for Whole Slide Image Classification with Self-supervised Contrastive...

Area of Science:

• Computational pathology
• Digital pathology
• Machine learning in medicine

Background:

• Whole slide images (WSIs) present classification challenges due to high resolution and lack of localized annotations.
• Multiple instance learning (MIL) is suitable for WSI classification when only slide-level labels are available.
• Existing MIL models struggle with large or unbalanced bags common in WSI data and memory constraints.

Purpose of the Study:

• To develop a novel MIL-based method for accurate WSI classification and tumor detection.
• To address the limitations of MIL models in handling large-scale WSI data and memory costs.
• To improve classification and localization accuracy through multiscale feature fusion.

Main Methods:

• A novel MIL aggregator employing a dual-stream architecture with trainable distance measurement.
• Self-supervised contrastive learning for robust feature extraction and mitigating memory issues with large bags.
• Pyramidal fusion mechanism for integrating multiscale WSI features.

Main Results:

• The proposed MIL method achieves classification accuracy comparable to fully-supervised approaches, with less than a 2% gap on WSI datasets.
• The model outperforms all previously reported MIL-based methods for WSI classification and tumor detection.
• The MIL aggregator demonstrates superior performance on general MIL problems, validated on standard MIL datasets.

Conclusions:

• The developed MIL framework effectively handles WSI classification and tumor detection challenges without requiring localized annotations.
• Self-supervised learning and pyramidal fusion significantly enhance MIL model performance and efficiency for high-resolution pathology images.
• This method offers a powerful and scalable solution for computational pathology, advancing diagnostic capabilities.