Hard example mining in Multi-Instance Learning for Whole-Slide Image Classification.

Area of Science:

• Computational pathology
• Digital pathology
• Machine learning for medical imaging

Background:

• Multiple Instance Learning (MIL) is effective for Whole-Slide Image (WSI) classification.
• Current MIL methods for WSIs use feature extractors and aggregators but struggle with hard-to-classify instances.
• Existing methods optimize based on easy instances, limiting discrimination of challenging cases.

Purpose of the Study:

• To develop a novel MIL method for WSI classification that addresses the limitations of existing approaches.
• To improve the discrimination of hard-to-classify instances in WSI analysis.
• Introduce Hard Pseudo-label Assignment (HPA-MIL) for direct mining of hard instances.

Main Methods:

• Implemented a novel Multiple Instance Learning (MIL) approach named HPA-MIL.
• Utilized pseudo-label assignment to directly mine hard-to-classify instances.
• Integrated feature extraction and MIL aggregation for interactive optimization.

Main Results:

• HPA-MIL achieved an Area Under the Curve (AUC) of 0.9523 on the TCGA NSCLC dataset.
• The proposed method demonstrated superior performance compared to existing state-of-the-art techniques.
• Effectively improved the classification of challenging instances in WSIs.

Conclusions:

• HPA-MIL offers a significant advancement in Multiple Instance Learning for WSI classification.
• The pseudo-label assignment strategy effectively targets and classifies hard instances.
• This method enhances the diagnostic capabilities of computational pathology tools.