Diffusion Probabilistic Learning With Gate-Fusion Transformer and Edge-Frequency Attention for Retinal Vessel Segmentation

Retinal vessel topology provides unique biological information for the diagnosis of fundus diseases. However, most existing deep learning-based vessel segmentation methods mainly focus on global fundus structure, which may suffer from generalization errors and blurring caused by lesions and image noise. Besides, vessel edge details and feature channel information are generally ignored or not considered simultaneously, and this insufficiency commonly leads to suboptimal segmentation performance. To tackle these issues, we propose a novel diffusion probabilistic learning with gate-fusion transformer and edge-frequency attention (DPL-GFT-EFA) for retinal vessel segmentation. Specifically, the DPL leverages the image denoising as a proxy task to pretrain the segmentation model, which enhances the anti-interference ability by learning noise-related information. Then, the gate-fusion transformer (GFT) block fuses high-level representations from condition and diffusion encoders (DEs) with a gate mechanism, highlighting the mutual features between fundus patterns and noisy images. Finally, the edge-frequency attention (EFA) block is introduced to further consolidate the vessel edge details and discriminative channel features. We conduct the experiments on five public retinal image datasets, and achieve the accuracies of 97.05%, 97.70%, 97.71%, 97.16%, and 97.26% on DRIVE, STARE, CHASE_DB1, HRF, and IOSTAR datasets, respectively. These results demonstrate that the proposed method outperforms state-of-the-art models and achieve promising segmentation performance even in complex images containing fundus lesions and noise. Our source code is available at https://github.com/YangLibuaa/DPL-GTF-EFA.