A multi-level decoder and multi-attention network for accurate retinal vessel segmentation

Retinal vessel segmentation approaches largely depend on neural networks with U-shaped structures. However, the skip connections in it will inevitably introduce background noise and irrelevant features, leading to the feature dilution. To address this issue, we propose MDG-Net, a novel Multi-level Decoder and Multi-Attention Feature Fusion Network for retinal vessel segmentation without relying on skip connections. Specifically, MDG-Net introduces a Multilevel-Decoder Structure (MDS) module, which replaces traditional skip connections by effectively utilizing low-level features and capturing multi-scale information during the decoding stage, thereby enhancing feature recovery. In addition, a Multi-Attention Feature Fusion (MAF) module is designed to expand the receptive field and improve semantic representation learning through attention mechanisms. The proposed MDG-Net was evaluated on five public retinal vessel datasets, including DRIVE, STARE, CHASE_DB1, IOSTAR, and LES-AV. Experimental results show that MDG-Net outperforms state-of-the-art methods, achieving higher accuracy and AUC scores across these datasets. MDG-Net provides a robust alternative to traditional U-shaped architectures by eliminating skip connections and reducing feature dilution. The source code is available at: https://github.com/123fengye741/MDG-Net .