Pseudo-Bag Mixup Augmentation for Multiple Instance Learning-Based Whole Slide Image Classification.

Area of Science:

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

Background:

• Whole Slide Image (WSI) classification commonly uses multiple instance learning (MIL).
• MIL training faces challenges like insufficient data and neural network memorization.
• These issues limit the performance and generalization of MIL models for WSI analysis.

Purpose of the Study:

• To introduce Pseudo-bag Mixup (PseMix), a novel data augmentation scheme for MIL-based WSI classification.
• To address data scarcity and sample memorization in MIL training.
• To improve the classification performance and robustness of MIL models on WSIs.

Main Methods:

• Developed Pseudo-bag Mixup (PseMix), a data augmentation strategy generalizing Mixup to WSIs using pseudo-bags.
• Ensured size and semantic alignment within the Mixup strategy for WSIs.
• Designed PseMix as an efficient, decoupled method independent of MIL model predictions.

Main Results:

• PseMix significantly improved classification performance for state-of-the-art MIL networks on WSIs.
• The method enhanced MIL model generalization in specific test scenarios.
• PseMix demonstrated increased robustness against patch occlusion and label noise.

Conclusions:

• PseMix is an effective data augmentation technique for MIL-based WSI classification.
• The proposed method overcomes key training limitations in MIL for WSIs.
• PseMix offers a valuable tool for advancing computational pathology and digital diagnostics.