A Graph-Based Multi-Scale Approach With Knowledge Distillation for WSI Classification.

Area of Science:

• Computational pathology
• Digital pathology
• Machine learning in healthcare

Background:

• Whole Slide Images (WSIs) are crucial for cancer diagnosis but are gigapixel-sized, making pixel-level annotation infeasible.
• Existing Multi Instance Learning (MIL) methods for WSI classification often ignore spatial instance correlations and single-scale resolutions.

Purpose of the Study:

• To develop an advanced MIL approach for WSI classification that addresses limitations of current methods.
• To leverage the full potential of multi-scale WSI data by incorporating spatial instance correlations.

Main Methods:

• Proposed DAS-MIL, a graph-based multi-scale MIL framework.
• Integrated a self-supervised feature extractor.
• Employed a graph-based architecture to model inter- and intra-scale spatial instance correlations.
• Utilized self-distillation loss between resolutions to bridge information gaps.

Main Results:

• DAS-MIL demonstrated superior performance on WSI classification tasks.
• Achieved a +2.7% AUC and +3.7% accuracy improvement on the Camelyon16 benchmark.
• Outperformed state-of-the-art (SOTA) methods in WSI classification.

Conclusions:

• The proposed graph-based multi-scale MIL approach effectively enhances WSI classification.
• DAS-MIL provides a more contextualized representation by considering spatial correlations and multi-scale information.
• This framework offers a promising direction for improving automated diagnosis in digital pathology.