CoMIL: A Contrastive CNN-Transformer Framework with Multi-Instance Learning for Whole-Slide Pathology Image...

Area of Science:

• Computational pathology
• Digital pathology
• Machine learning for medical imaging

Background:

• Whole slide image (WSI) classification presents challenges due to gigapixel scale and weak supervision.
• Existing methods struggle to balance global context with local details in WSI analysis.
• Weak labels in datasets can lead to reduced model robustness.

Purpose of the Study:

• To propose CoMIL, a novel dual-branch framework for improved WSI classification.
• To address the limitations of single-stream networks in capturing both global and local information.
• To enhance spatial awareness and model robustness against label noise in WSI data.

Main Methods:

• CoMIL utilizes a dual-branch architecture with a Transformer branch for long-range dependencies and a CNN branch for local morphology.
• A Hyper Positional Generator (HyperPG) module is introduced to mitigate spatial information loss using multi-scale adaptive mechanisms and deformable convolutions.
• Symmetric mutual learning with KL divergence minimization is employed to improve robustness against weak label noise.

Main Results:

• CoMIL achieved an area under the curve (AUC) of 98.6% and 95.3% accuracy on the Camelyon16 dataset.
• The method reached an AUC of 98.8% and 93.3% accuracy on the TCGA_Kidney dataset.
• Performance surpassed known advanced WSI classification methods on both datasets.

Conclusions:

• The proposed CoMIL framework effectively balances global context and local details for WSI classification.
• HyperPG module enhances spatial awareness with linear complexity, improving WSI analysis.
• Symmetric mutual learning enhances model robustness, making it suitable for weakly supervised WSI classification tasks.