Multi-Scale Efficient Graph-Transformer for Whole Slide Image Classification.

Area of Science:

• Digital pathology
• Computational oncology
• Medical image analysis

Background:

• Multi-scale information is crucial for accurate cancer diagnosis from whole slide images (WSIs).
• Existing multi-scale vision Transformers struggle with the immense size of gigapixel WSIs.
• Efficiently processing high-resolution medical imaging data remains a challenge.

Purpose of the Study:

• To introduce a novel Multi-scale Efficient Graph-Transformer (MEGT) framework for whole slide image (WSI) classification.
• To address the limitations of current methods in handling gigapixel WSIs.
• To improve the accuracy and efficiency of cancer diagnosis using large-scale histopathological data.

Main Methods:

• Proposed a Multi-scale Efficient Graph-Transformer (MEGT) framework utilizing two independent Efficient Graph-based Transformer (EGT) branches for low and high-resolution WSI patch embeddings.
• Integrated graph representation into Transformers within EGT to capture spatial relationships and local-global information.
• Developed a novel Multi-scale Feature Fusion Module (MFFM) with cross-attention to bridge semantic gaps between different resolution features.
• Implemented a token pruning module in EGT to reduce redundant tokens and accelerate training.

Main Results:

• The MEGT framework demonstrated significant effectiveness in WSI classification tasks.
• Experiments on TCGA-RCC and CAMELYON16 datasets validated the proposed model's performance.
• The dual-branch approach and MFFM successfully handled multi-scale information in gigapixel WSIs.

Conclusions:

• The proposed MEGT framework offers an effective solution for cancer diagnosis using large-scale whole slide images.
• MEGT overcomes the computational challenges associated with gigapixel WSI analysis.
• This approach advances the application of deep learning in digital pathology and computational oncology.