<p>Accurate segmentation of retinal blood vessels in the fundus is a crucial step in the auxiliary diagnosis of diabetic retinopathy. To address issues such as poor localization throughout the process, vessel discontinuities, and capillary loss in retinal fundus image vascular segmentation, we propose a new method called WFL-Vnet. This method adopts a V-shaped architecture, utilizing ConvNeXt as the encoder and U-Net as the decoder, with the integration of the Res-CA (Residual Coordinate Attention) module to enhance feature localization. In the encoder stage, we select ConvNeXt blocks with large convolutional kernels at an appropriate ratio to improve the model’s global perception of the fundus vasculature. Additionally, the Res-CA module is incorporated into each convolutional layer to help locate fine vessels, thereby enhancing the model’s ability to pinpoint tiny blood vessels at the pixel level. Next, we present a segmentation approach that combines ConvNeXt as the encoder and U-Net as the decoder, which not only aids in feature localization but also strengthens the model’s capacity to detect small vessels at a granular level. Furthermore, the Parallel ConvNeXt block is introduced in the fourth feature extraction layer of the encoder to prevent the loss of crucial blood vessel feature information during the deep feature map processing of the segmentation task.In the decoder stage, the U-Net structure is employed, with the Res-CA module added to each layer to improve the model’s ability to decode and predict vessel features across the entire segmentation process.The proposed method was evaluated on three benchmark datasets: DRIVE, CHASE_DB1, and STARE, achieving the following results: F1-scores of 84.74%, 83.64%, and 84.54%, Accuracy (ACC) of 97.40%, 97.55%, and 97.86%, and Specificity (SPE) of 98.83%, 98.95%, and 99.34%, respectively. These experimental results demonstrate that the WFL-Vnet method outperforms existing mainstream segmentation techniques.</p>