Unsupervised domain adaptation multi-level adversarial learning-based crossing-domain retinal vessel segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Biomedical Engineering

Background:

• Retinal vasculature segmentation is vital for diagnosing diseases like glaucoma and retinopathy.
• Current models struggle with cross-source fundus images, limiting diagnostic accuracy.
• Accurate segmentation is essential for understanding and diagnosing various systemic and ocular diseases.

Purpose of the Study:

• To develop a robust framework for accurate retinal vessel segmentation across diverse fundus image sources.
• To overcome the limitations of existing models in handling cross-domain segmentation challenges.
• To enhance the precision of retinal vessel segmentation for improved disease diagnosis.

Main Methods:

• Proposed a novel Multi-level Adversarial Learning and Pseudo-label Denoising-based Self-training Framework (MLAL&PDSF).
• Employed multi-level adversarial networks at feature and image layers to align source and target domain distributions.
• Utilized distance comparison for refining pseudo-labels and correcting inaccuracies in self-training.

Main Results:

• Achieved remarkable unsupervised domain adaptive segmentation performance on multiple datasets (CHASEDB1, STARE, HRF).
• Demonstrated high AUC, sensitivity, specificity, accuracy, and F1-scores in cross-domain segmentation tasks (e.g., DRIVE to CHASEDB1 and STARE).
• Validated the framework's efficacy through extensive comparative experiments.

Conclusions:

• MLAL&PDSF effectively achieves accurate segmentation for cross-domain retinal vessel datasets.
• The framework provides a robust foundation for advancing cross-domain segmentation techniques.
• Enhanced segmentation accuracy aids in better diagnosis and understanding of retinal diseases.