AFFD-Net: A Dual-Decoder Network Based on Attention-Enhancing and Feature Fusion for Retinal Vessel Segmentation

Abstract

The morphological characteristics of retinal vessels in the fundus serve as the primary basis for diagnosing and assessing the risk of ophthalmic diseases. An effective segmentation scheme for retinal vessels can aid in the early diagnosis and treatment of these diseases, as well as help prevent their progression. However, accurate vessel segmentation is challenging due to the low contrast of fundus images and the complexity of the vessels’ morphological structure. To address the low sensitivity and poor generalization ability of the existing methods in vascular extraction, a Dual-decoder Network based on Attention-enhancing and multi-scale Feature Fusion (AFFD-Net) is proposed. AFFD-Net inherits the codec concept of U-Net. To improve the performance of our U-Net model, we made two modifications. Firstly, we reduced the number of convolution kernel filters in each layer, thereby significantly reducing the number of training parameters. This helps to avoid overfitting and improves the model’s ability to generalize. Secondly, we added a Multi-scale Feature Extraction (MFE) module and an M/A intermediate decoder to enhance the model’s sensitivity. MFE is designed as the first encoding unit of AFFD-Net to obtain rich vascular features in the complex anatomical background and adapt to the large-scale variations of vessels. The M/A intermediate decoder is composed of the Multi-scale Feature Fusion (MFF) module and the Attention-enhancing Hybrid Feature Fusion (AHFF) module. The MFF module integrates deep semantic information and shallow spatial information to ensure that the features at each scale in the middle layer are fully utilized. The AHFF module adaptively fuses the hybrid features at different scales to generate two feature descriptors with different focuses which can improve the expressiveness of the model. AFFD-Net is evaluated on three public databases including DRIVE, STARE, and CHASE_DB1, and the sensitivity values obtained are 84.19%, 84.58%, and 82.62%, respectively. It has higher sensitivity and better generalization ability than other state-of-the-art methods. Compared with classical networks including U-Net, U-Net++, and U-Net3+, AFFD-Net has fewer parameters and higher segmentation accuracy. Our proposed segmentation model exhibits superior performance across a range of metrics, indicating its promising potential for practical applications.