Learnable Context in Multiple Instance Learning for Whole Slide Image Classification and Segmentation.

Area of Science:

• Digital pathology
• Computational biology
• Machine learning

Background:

• Multiple instance learning (MIL) is crucial for whole slide image (WSI) analysis, treating WSIs as bags of instances.
• Current MIL methods often overlook the contextual relationships between instances, potentially limiting performance.

Purpose of the Study:

• To enhance instance representations by learning contextual features between instances.
• To improve feature aggregation in MIL for WSI analysis, especially in cases with sparse positive instances.
• To develop a more robust and accurate MIL framework for WSI classification and segmentation.

Main Methods:

• Proposed a novel approach that learns contextual features between instances to enrich instance representations.
• Introduced a self-attention mechanism for feature aggregation to better capture instance correlations.
• Evaluated the method on Camelyon16 and TCGA-NSCLC datasets for WSI classification and segmentation tasks.

Main Results:

• Achieved 1-4% higher classification accuracy compared to existing WSI classification methods on Camelyon16 and TCGA-NSCLC datasets.
• Outperformed the latest weakly supervised WSI segmentation method by 0.6 in Dice coefficient on the Camelyon16 dataset.
• Demonstrated the effectiveness of incorporating instance context and self-attention for improved WSI analysis.

Conclusions:

• The proposed method significantly improves WSI classification and segmentation accuracy by leveraging instance context and self-attention.
• This approach offers a more robust solution for analyzing WSIs, particularly in challenging scenarios with limited positive instances.
• The findings highlight the potential of contextual MIL for advancing digital pathology and computational analysis of histopathological images.