Joint semi-supervised and contrastive learning enables domain generalization and multi-domain segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Deep learning models struggle with domain shifts in medical images.
• Variations in appearance and content across datasets hinder model generalization.
• Accurate segmentation of retinal Optical Coherence Tomography (OCT) images is crucial for diagnosing eye diseases.

Purpose of the Study:

• To introduce SegCLR, a versatile framework for robust image segmentation across different domains.
• To leverage both supervised and contrastive learning for improved performance with labeled and unlabeled data.
• To enhance the generalizability of deep learning models for multi-domain medical image analysis.

Main Methods:

• SegCLR framework combines supervised and contrastive learning.
• Evaluated on three diverse 3D retinal OCT datasets for fluid segmentation.
• Tested across various network configurations and initializations.
• Proposed a domain generalization extension for zero-shot adaptation.

Main Results:

• SegCLR achieved performance comparable to supervised models trained on the target domain in unsupervised domain adaptation.
• Segmentation performance showed minimal impact from the amount of unlabeled target domain data.
• The domain generalization extension enabled effective segmentation without target domain information.
• Models demonstrated superior generalizability to both in- and out-of-domain test data.

Conclusions:

• The addition of contrastive loss to supervised segmentation training yields inherently more generalizable models.
• SegCLR offers a pragmatic solution for multi-domain segmentation, adaptable to varying data availability (labeled, unlabeled, or none).
• The framework advances deep learning-based segmentation for multi-domain applications in medical imaging.