D2PF-net: A dual-decoder progressive fusion network for retinal vessel segmentation

Zirun Zhang, Jingqin Lv, Yafeng Ren, Yuejin Zhang, Chengxin Ye, Nan Jiang, Donghong Ji, Qingzhong Zhou, Hongbin ZhangBiomedical Signal Processing and Control 2026 · OpenReview

Abstract

Accurate but efficient retinal vessel segmentation is crucial for clinical diagnosis. Although current deep learning-based methods have achieved satisfactory segmentation performance, there are still some challenges, such as the lack of microvessels and insufficient global modelling. To address these problems, we propose a Dual-Decoder Progressive Fusion Network (D2PF-Net). First, a multi-scale vessel enhancement (MVE) module is devised to alleviate the lack of microvessels. Second, a decoding feature fusion (DFF) module is embedded into the left branch of D2PF-Net to progressively enrich the decoder with sufficient global contextual information. Simultaneously, a cross-scale context fusion (CCF) module is inserted into the right branch of D2PF-Net to make the encoder focus more on multi-scale pathological information. Finally, a multi-scale feature aggregation (MFA) module is designed to aggregate complementary multi-scale semantics information from different branches. Extensive experiments have been conducted on five retinal vessel and two polyp image datasets, validating the efficiency and generalization ability of D2PF-Net, which is quite competitive compared to the state-of-the-art methods. More importantly, D2PF-Net is relatively lightweight, efficient and interpretable for real-time prediction, further promoting its practicality.
BibTeX
@article{zhang2026d,
  author = {Zirun Zhang and Jingqin Lv and Yafeng Ren and Yuejin Zhang and Chengxin Ye and Nan Jiang and Donghong Ji and Qingzhong Zhou and Hongbin Zhang},
  title = {{D2PF-net: A dual-decoder progressive fusion network for retinal vessel segmentation}},
  journal = {Biomedical Signal Processing and Control},
  year = {2026},
  url = {https://openreview.net/forum?id=9JDA7hmTVD},
}

Loading related content...

Figure