An efficient context-aware approach for whole-slide image classification.

Area of Science:

• Digital pathology
• Computational pathology
• Artificial intelligence in medicine

Background:

• Accurate disease diagnosis from gigapixel whole-slide images (WSIs) is crucial but challenging.
• Existing computational pathology methods struggle with holistic analysis of WSI patch dependencies.

Purpose of the Study:

• To introduce WSI inspection via transformer (WIT), a novel context-aware approach for slide-level classification.
• To improve the accuracy and interpretability of computational pathology for disease diagnosis.

Main Methods:

• Developed WIT, a transformer-based model that aggregates patch features for WSI representation.
• Employed a context-aware strategy to model dependencies among patches within WSIs.
• Evaluated WIT's performance against state-of-the-art baseline methods on multiple cancer datasets.

Main Results:

• WIT achieved high accuracy in detecting 32 cancer types (82.1%) on TCGA and diagnosing cancer (0.918) on CPTAC.
• Demonstrated superior performance in prostate cancer diagnosis (0.882) from needle biopsy slides, outperforming baselines by significant margins.
• WIT successfully identified influential WSI regions contributing to classification decisions.

Conclusions:

• WIT represents a new paradigm in computational pathology, offering enhanced accuracy and interpretability.
• The approach facilitates the development of advanced digital pathology tools for disease diagnosis.
• Context-aware patch aggregation in WIT significantly improves slide-level classification performance.