DEAF-Net: Detail-Enhanced Attention Feature Fusion Network for Retinal Vessel Segmentation

Area of Science:

• Medical Imaging
• Computer Vision
• Biomedical Engineering

Background:

• Retinal vessel segmentation is vital for diagnosing ophthalmic and cardiovascular diseases.
• Challenges include dense, irregular vessels, low contrast, and feature loss in encoder-decoder networks.
• Existing attention mechanisms overlook multidimensional feature importance.

Purpose of the Study:

• To propose a novel Detail-Enhanced Attention Feature Fusion Network (DEAF-Net) for improved retinal vessel segmentation.
• To address limitations of existing methods in preserving fine details and integrating multidimensional features.
• To enhance the accuracy of segmenting delicate retinal vasculature.

Main Methods:

• Developed a Detail-Enhanced Residual Block (DERB) to maintain intricate vessel features.
• Introduced a Multidimensional Collaborative Attention Encoder (MCAE) for optimized information extraction.
• Utilized a Dynamic Decoder (DYD) to minimize information loss during upsampling.
• Integrated these modules into a DEFF module for multi-scale feature fusion.

Main Results:

• DEAF-Net achieved high performance on DRIVE, CHASEDB1, and STARE datasets.
• Sensitivity scores reached up to 0.8784, and AUC scores up to 0.9913.
• Demonstrated superior segmentation of fine retinal vessels compared to existing methods.

Conclusions:

• DEAF-Net effectively overcomes limitations in current retinal vessel segmentation techniques.
• The proposed network enhances detail preservation and multidimensional feature integration.
• DEAF-Net shows significant potential for clinical applications in diagnosing eye and cardiovascular diseases.