DGR-MIL: Exploring Diverse Global Representation in Multiple Instance Learning for Whole Slide Image Classification

Area of Science:

• Computational pathology
• Machine learning
• Weakly supervised learning

Background:

• Multiple Instance Learning (MIL) is crucial for whole slide image (WSI) classification in tumor detection.
• Current MIL methods primarily model instance correlation, neglecting instance diversity, leading to suboptimal performance and high computational costs.

Purpose of the Study:

• To propose a novel MIL aggregation method, Diverse Global Representation MIL (DGR-MIL), that effectively models instance diversity.
• To improve the accuracy and efficiency of tumorous lesion detection in WSIs using MIL.

Main Methods:

• Developed DGR-MIL, an aggregation method using global vectors to represent instance diversity.
• Employed a cross-attention mechanism to model instance-global vector similarity, reflecting instance correlation.
• Introduced positive instance alignment and a determinantal point process-based diversification learning paradigm to enhance global vector representation.

Main Results:

• DGR-MIL significantly outperforms state-of-the-art MIL aggregation models.
• Achieved superior performance on the CAMELYON-16 and TCGA-lung cancer datasets.
• Demonstrated improved modeling of instance diversity for better WSI classification.

Conclusions:

• DGR-MIL offers a powerful new approach for weakly supervised learning in computational pathology.
• The method effectively captures instance diversity, leading to enhanced tumor detection in WSIs.
• The proposed techniques provide efficient and theoretically sound diversification learning for MIL.