Enhanced feature dynamic fusion gated UNet for robust retinal vessel segmentation

TL;DR

This paper introduces a new deep learning model for accurately segmenting retinal vessels in images, especially in challenging areas like small vessels and lesions.

Contribution

The novel EFDG-UNet model integrates dynamic fusion, global position modeling, and adaptive attention for improved retinal vessel segmentation.

Findings

EFDG-UNet achieved an AUC of 0.9932 and F1-score of 0.8469 on the CHASE_DB1 dataset.

The model outperformed baseline methods in low-contrast and complex vessel regions.

It showed strong performance across multiple datasets including DRIVE and STARE.

Abstract

This study proposes a Deep learning model, the Enhanced Feature Dynamic Fusion-Gated U-Net (EFDG-UNet), for retinal vessel segmentation. To address challenges in segmenting small vessels, handling lesion interference, and adapting to multi-scale structures, the model incorporates optimized feature fusion, dynamic selection, and global position modeling. The Feature Navigation Hub (FN-Hub) captures long-range dependencies across multiple encoder layers, improving multi-scale vessel segmentation. The Adaptive Gated Residual Block (AGRB) uses a dynamic gating mechanism to enhance feature selectivity in lesion areas and low-contrast scenarios. The Parallel Focused Attention Module (PFAM) optimizes channel and spatial information for fine-grained vessel features. Experimental validation on DRIVE, CHASE_DB1, and STARE datasets shows that EFDG-UNet achieves state-of-the-art performance, attaining an AUC of 0.9932 and F1-score of 0.8469 on CHASE_DB1, and an AUC of 0.9886 and F1-score of 0.8412 on DRIVE. The model shows improved performance in low-contrast regions and complex vessel structures compared to baseline methods.

Figures