Self-rule to multi-adapt: Generalized multi-source feature learning using unsupervised domain adaptation for...

Area of Science:

• Digital Pathology
• Machine Learning
• Computer Vision

Background:

• Supervised learning in digital pathology is limited by the high cost of acquiring labeled data.
• Pre-trained models often struggle to generalize across different tissue staining, types, and textures.
• Existing domain adaptation methods typically require fully-labeled source datasets.

Purpose of the Study:

• To develop a domain adaptation method that does not require fully-labeled source datasets.
• To enable effective knowledge transfer from limited labeled source data to new target domains in digital pathology.
• To leverage self-supervised learning for robust domain adaptation.

Main Methods:

• Proposed Self-Rule to Multi-Adapt (SRMA) framework utilizing self-supervised learning for domain adaptation.
• Employed intra-domain and cross-domain self-supervision to capture visual similarities.
• Developed a generalized formulation for learning from multiple source domains.

Main Results:

• SRMA demonstrated superior performance in colorectal tissue type classification compared to baseline methods in both single and multi-source domain adaptation settings.
• The method effectively transferred discriminative knowledge without requiring additional annotations in the target domain.
• Validated the approach on an in-house clinical cohort, confirming its practical applicability.

Conclusions:

• SRMA offers a powerful solution for domain adaptation in digital pathology, overcoming the limitations of labeled data scarcity.
• The self-supervised approach enhances model generalization across diverse tissue domains.
• The open-source availability of code and models facilitates further research and application.