A feature aggregation and feature fusion network for retinal vessel segmentation

• Published in: Biomedical signal processing and control Vol. 85; p. 104829
• Main Authors: Ni, Jiajia, Sun, Haizhou, Xu, Jinxin, Liu, Jinhui, Chen, Zhengming
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.08.2023
• Subjects: Attention mechanisms; Feature fusion; Multi-scale feature; Neural networks; Vessel segmentation; Vessel segmentation; Attention mechanisms; Feature fusion; Neural networks; Multi-scale feature
• ISSN: 1746-8094; 1746-8108
• DOI: 10.1016/j.bspc.2023.104829

•The multi-scale feature aggregation (MFA) block is proposed to effectively extract multi-scale context information, which can improve the segmentation ability of the proposed network.•The feature reuse and distribution (FRD) block is proposed to adapt high-level features information and spatial information.•The attention feature fusion (AFF) block is introduced to combine high-level features and shadow features, which can not only effectively decrease redundant background noise information in high-level features but also retain spatial information in shadow features. Neural networks have achieved outstanding performance in retinal vessel segmentation. However, since its continuous upsampling and convolution operation in the decoding stage, the semantic information and class information of the high-level features are destroyed. To address these problems, we proposed a new feature aggregation and feature fusion network (FAF-Net). Firstly, we introduced a multi-scale feature aggregation (MFA) block, which adjusts the receptive fields to learn more multi-scale features information. Furthermore, a feature reuse and distribution (FRD) block is intended to preserve the multi-scale feature information of the image and reduce the background noises in the feature map. Finally, the attention feature fusion (AFF) block is employed to effectively reduce the information loss of high-level features and connect the encoding and decoding stages. This multi-path combination helps to learn better representations and more accurate vessel feature maps. We evaluate the network on three retinal image databases (DRIVE, CHASEDB1, STARE). The proposed network outperforms existing current state-of-the-art vessel segmentation methods. Comprehensive experiments prove that FAF-Net is suited to processing medical image segmentation with limited samples and complicated features.