Self-Supervised Representation Distribution Learning for Reliable Data Augmentation in Histopathology WSI...

Area of Science:

• Computational pathology
• Digital pathology
• Machine learning in healthcare

Background:

• Whole slide image (WSI) classification relies on patch-level representations from pre-trained encoders.
• Current methods often freeze these representations, limiting classifier robustness due to invariant features.
• This invariance hinders the diverse representation needs for effective MIL classifier training.

Purpose of the Study:

• To introduce a novel framework, Self-Supervised Representation Distribution Learning (SSRDL), for improved patch-level representation learning in WSI classification.
• To enhance WSI-level data augmentation through an online representation sampling strategy (ORS).
• To address the limitations of invariant representations in current Multiple Instance Learning (MIL) approaches.

Main Methods:

• Proposed the Self-Supervised Representation Distribution Learning (SSRDL) framework.
• Incorporated an online representation sampling strategy (ORS) for both patch feature extraction and WSI data augmentation.
• Evaluated the method on three distinct datasets using three different MIL frameworks.

Main Results:

• SSRDL demonstrated superior performance in histopathology image representation learning.
• The framework significantly improved data augmentation for WSI analysis.
• Outperformed state-of-the-art methods across various WSI classification frameworks.

Conclusions:

• SSRDL effectively enhances patch-level representations and WSI data augmentation.
• The proposed method offers a robust solution for MIL-based WSI classification.
• Achieved state-of-the-art results, indicating the potential of dynamic representation learning.