Attention2majority: Weak multiple instance learning for regenerative kidney grading on whole slide images.

Area of Science:

• Digital pathology
• Computational biology
• Machine learning in medicine

Background:

• Deep learning excels in medical image analysis (CT, PET, MRI) but struggles with large whole slide images (WSIs).
• WSIs present computational challenges due to their massive size, hindering efficient deep learning processing.
• Automated WSI classification is crucial for accurate diagnosis and research.

Purpose of the Study:

• To develop an efficient and accurate deep learning model for classifying whole slide images (WSIs).
• To introduce a novel weak multiple instance learning approach, attention2majority, for automated WSI analysis.
• To enhance slide-level classification performance and interpretability in digital pathology.

Main Methods:

• Attention2majority employs a weak multiple instance learning framework for WSI classification.
• It trains a CNN on label-free patches, followed by intelligent sampling of high-confidence patches.
• A multi-head attention model performs slide-level classification using these intelligently sampled patches.

Main Results:

• Attention2majority achieved an average AUC of 97.4%±2.4 on classifying regenerated kidney sections (127 WSIs).
• The intelligent sampling module outperformed random sampling, boosting performance from 94.9%±3.1 to 97.4%±2.4 AUC.
• A variation on the Camelyon16 dataset yielded an AUC of 89.1%±0.8, demonstrating broad applicability.

Conclusions:

• Attention2majority offers an efficient and accurate framework for automated WSI classification.
• Intelligent patch sampling significantly enhances model performance and interpretability over random sampling.
• The model provides excellent slide-level interpretability and enables multi-class predictions for WSIs.