TCDDU-Net: combining transformer and convolutional dual-path decoding U-Net for retinal vessel segmentation.

Area of Science:

• Medical imaging analysis
• Deep learning for ophthalmology

Background:

• Accurate retinal blood vessel segmentation is vital for diagnosing eye diseases.
• Challenges include small vessel size, complex structures, and low contrast in fundus images.

Purpose of the Study:

• To develop an advanced deep learning model for improved retinal vessel segmentation.
• To address the limitations of existing methods in segmenting complex retinal vasculature.

Main Methods:

• Proposed TCDDU-Net, a dual-path U-Net incorporating transformer and convolutional components.
• Introduced a selective dense connection Swin transformer block for feature fusion and long-distance dependency capture.
• Designed a background decoder using deformable convolution for enhanced segmentation.

Main Results:

• Achieved high segmentation accuracies of 96.98% (DRIVE), 97.40% (STARE), and 97.23% (CHASE).
• Obtained excellent AUC metrics of 98.68% (DRIVE), 98.56% (STARE), and 98.50% (CHASE).
• Demonstrated superior performance over existing methods across multiple datasets using F1 score, specificity, and sensitivity.

Conclusions:

• The TCDDU-Net method significantly enhances retinal vessel segmentation performance.
• The proposed approach offers a robust solution for clinical applications in ophthalmology.
• This advancement contributes to improved diagnostic efficiency and disease management.