SFA-Net: Scale and Feature Aggregate Network for Retinal Vessel Segmentation.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Deep Learning

Background:

• U-Net based networks show promise in retinal vessel segmentation.
• Previous methods often overlook shallow-level features, potentially limiting accuracy.
• Upsampling and convolution operations can degrade semantic information in decoder layers.

Purpose of the Study:

• To propose a novel Scale and Feature Aggregate Network (SFA-Net) for improved retinal vessel segmentation.
• To leverage both multi-scale high-level and shallow-level features effectively.
• To address information loss during feature fusion in deep learning models.

Main Methods:

• Embedded a residual atrous spatial feature aggregate (RASF) block in the encoder to capture multiscale information.
• Introduced an attentional feature module (AFF) for enhanced fusion of shallow and high-level features.
• Designed a multi-path feature fusion (MPF) block to integrate decoder layer features and mitigate information loss.

Main Results:

• SFA-Net demonstrated competitive performance on DRIVE, STARE, and CHASE_DB1 datasets.
• The network effectively utilizes multiscale high-level and shallow features.
• Experimental results confirm the efficacy of the proposed modules (RASF, AFF, MPF).

Conclusions:

• SFA-Net offers an effective solution for retinal vessel segmentation.
• The proposed architecture is suitable for analyzing complex medical images.
• Integrating shallow and multi-level features is crucial for enhancing segmentation accuracy.