Disentangled Pseudo-Bag Augmentation for Whole Slide Image Multiple Instance Learning.

Area of Science:

• Computational pathology
• Artificial intelligence in medicine
• Digital pathology

Background:

• Multiple instance learning (MIL) is standard for whole slide image (WSI) classification.
• Existing MIL methods face challenges due to limited labeled WSIs and noisy pseudo-bag augmentation.
• Improving data augmentation is crucial for enhancing MIL performance in digital pathology.

Purpose of the Study:

• To introduce DPBAug, a novel pseudo-bag generation paradigm for WSI data augmentation.
• To address limitations of noisy pseudo-labels and low-quality pseudo-bags in current MIL augmentation methods.
• To improve the classification performance, reliability, and data efficiency of MIL methods for WSI analysis.

Main Methods:

• DPBAug employs an intra-slide pseudo-bag generation module with phenotype partitioning and instance sampling with replacement.
• An inter-slide pseudo-bag fusion module integrates information across multiple WSIs for high-quality sample generation.
• A pseudo-bag memory update module prioritizes valuable synthetic pseudo-bags to boost network performance.

Main Results:

• DPBAug significantly outperforms existing augmentation methods on various public WSI datasets.
• The proposed method enhances the classification performance and reliability of multiple MIL baselines.
• DPBAug demonstrates improved generalization and data efficiency for MIL methods.

Conclusions:

• DPBAug offers a robust solution for WSI data augmentation in MIL.
• The method facilitates the adoption of MIL in clinical practice and rare cancer research.
• DPBAug advances the field of computational pathology by improving WSI classification capabilities.