Cross-patch feature interactive net with edge refinement for retinal vessel segmentation

Area of Science:

• Medical image analysis
• Deep learning for biomedical imaging

Background:

• Accurate retinal vessel segmentation is crucial for diagnosing eye diseases.
• Existing deep learning methods struggle with low contrast and thin vessels, leading to segmentation errors and loss of detail.

Purpose of the Study:

• To develop an advanced deep learning model for precise end-to-end retinal vessel segmentation.
• To improve the segmentation of thin vessels and vessel edges, ensuring continuity and integrity of the vessel skeleton.

Main Methods:

• Proposed a Cross-patch Feature Interactive Net (CFI-Net) with a dual-decoder architecture.
• Introduced a Joint Refinement Down-Sampling Method (JRDM) to preserve feature information during encoding.
• Developed a Cross-patch Interactive Attention Mechanism (CIAM) and Adaptive Spatial Context Guide Method (ASCGM) for enhanced feature interaction and detail preservation.

Main Results:

• CFI-Net demonstrated superior performance on retinal and coronary angiography datasets.
• Achieved outstanding results in comprehensive segmentation metrics like AUC and CAL.
• Significantly improved segmentation of thin vessels and vessel edges compared to existing methods.

Conclusions:

• The proposed CFI-Net effectively addresses limitations in current retinal vessel segmentation techniques.
• The model enhances segmentation accuracy, continuity, and detail, particularly for challenging cases.
• CFI-Net offers a robust solution for clinical diagnosis support through improved medical image analysis.