Shapley Values-Enabled Progressive Pseudo Bag Augmentation for Whole-Slide Image Classification.

Area of Science:

• Computational pathology
• Digital pathology
• Machine learning in medicine

Background:

• Whole-slide image (WSI) classification is challenging due to gigapixel resolution and limited annotations.
• Multiple-instance learning (MIL) is a weakly supervised method for WSI, but refining instance-level information from bag-level labels is difficult.
• Conventional MIL methods often suffer from skewed attention distributions and inaccurate instance identification.

Purpose of the Study:

• To develop an improved method for instance importance scoring (IIS) in MIL for WSI classification.
• To enhance the accuracy and interpretability of MIL models in computational pathology.
• To address the limitations of traditional attention-based IIS estimation.

Main Methods:

• Proposed a novel approach inspired by cooperative game theory, utilizing Shapley values to assess instance contributions for improved IIS estimation.
• Accelerated Shapley value computation using attention mechanisms while maintaining enhanced instance identification and prioritization.
• Introduced a framework for progressive pseudo-bag assignment based on estimated IIS to promote balanced attention distributions.

Main Results:

• Demonstrated superior performance over state-of-the-art methods on CAMELYON-16, BRACS, TCGA-LUNG, and TCGA-BRCA datasets.
• Achieved enhanced interpretability and class-wise insights in WSI classification.
• Validated the effectiveness of Shapley values and pseudo-bag assignment for MIL in computational pathology.

Conclusions:

• The proposed Shapley value-based MIL approach significantly improves WSI classification accuracy and interpretability.
• The method effectively refines instance-level information, overcoming limitations of conventional attention mechanisms.
• This work offers a promising direction for advancing weakly supervised learning in digital pathology.