Self-supervised Learning through Multi-magnification Feature Correspondence for Histopathological Image Analysis.

Area of Science:

• Digital Pathology
• Machine Learning
• Medical Imaging

Background:

• Deep learning shows promise in pathological imaging analysis, but acquiring large labeled datasets is challenging due to the need for expert annotation.
• Self-supervised learning (SSL) offers a solution by utilizing unlabeled pathological images, with models pre-trained on relevant domains outperforming those trained on natural images.

Purpose of the Study:

• To develop a novel SSL method for pathological imaging that extracts features reflecting diagnostic processes across multiple structures and magnifications.
• To enhance feature representation consistency across different scales while preserving scale-specific characteristics.

Main Methods:

• Proposed a self-supervised learning approach that aligns feature representations with microscopic examinations integrating multi-magnification observations.
• Ensured similarity of feature representations for the same tissue region across different magnifications.
• Maintained distinctive characteristics observed at each scale during feature learning.

Main Results:

• The proposed SSL method demonstrated superior performance compared to previous SSL methods in four pathological image classification tasks.
• The method effectively learned consistent feature representations across various structures and scales relevant to pathological diagnosis.

Conclusions:

• The novel SSL method enhances pathological image analysis by effectively integrating information from multiple magnifications.
• This approach improves diagnostic performance by capturing a more comprehensive understanding of tissue architecture and cellular details.