ProDiv: Prototype-driven consistent pseudo-bag division for whole-slide image classification.

Area of Science:

• Computational pathology
• Digital pathology image analysis
• Machine learning in healthcare

Background:

• Pathology image classification is crucial for cancer diagnosis.
• Whole-Slide Images (WSIs) with weak labels pose challenges for traditional methods.
• Pseudo-bag-based multiple instance learning (MIL) is a promising approach for WSI classification.

Purpose of the Study:

• To propose an improved method for dividing pseudo-bags in MIL frameworks for pathology images.
• To enhance the classification performance of existing MIL methods by optimizing pseudo-bag generation.
• To address the limitations of random or clustering-based pseudo-bag division schemes.

Main Methods:

• A Prototype-driven Division (ProDiv) scheme is introduced for WSI pseudo-bag generation.
• An attention-based method generates a 'bag prototype' for each slide.
• WSI patch instances are clustered based on feature similarity to the prototype, forming pseudo-bags by combining instances from different clusters.

Main Results:

• The ProDiv scheme, integrated with existing MIL methods, achieved significant improvements in classification performance.
• AUC improvements of up to 7.3% and 10.3% were observed on two public WSI datasets.
• The effectiveness of ProDiv was validated through empirical results and experimental visualization.

Conclusions:

• The ProDiv scheme consistently improves the performance of MIL models in pathology image classification.
• The proposed method offers a practical and effective approach to pseudo-bag division for MIL frameworks.
• ProDiv demonstrates the potential to advance automated cancer diagnosis through improved WSI analysis.