Retinal Vessel Segmentation With Skeletal Prior and Contrastive Loss

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Retinal vessel morphology is crucial for diagnosing ophthalmic diseases.
• Deep learning has advanced retinal vessel segmentation, but challenges persist.
• Obstructions like optic discs/lesions and thin vessels are difficult to segment accurately; datasets are often small.

Purpose of the Study:

• To introduce SkelCon, a novel network for improved retinal vessel segmentation.
• To address limitations in segmenting thin vessels and vessels obscured by retinal structures.
• To overcome challenges posed by limited labeled fundus image datasets.

Main Methods:

• Proposed SkelCon network incorporating skeletal prior and contrastive loss.
• Developed a skeleton fitting module to maintain vessel morphology and improve thin vessel continuity.
• Employed contrastive loss for better vessel-background discrimination and data augmentation for robustness.

Main Results:

• SkelCon achieved state-of-the-art performance across multiple datasets (DRIVE, STARE, CHASE, HRF, UoA-DR, IOSTAR, RC-SLO, RFMiD, JSIEC39).
• Significantly outperformed existing methods in extracting thin vessels near lesions and optic discs.
• Demonstrated improved connectivity, overlapping area, vessel length consistency, sensitivity, specificity, and accuracy.

Conclusions:

• SkelCon effectively enhances retinal vessel segmentation, particularly for challenging cases.
• The proposed methods address key limitations in current deep learning-based segmentation techniques.
• The model shows promise for clinical applications in ophthalmic disease diagnosis.