LEA U-Net: a U-Net-based deep learning framework with local feature enhancement and attention for retinal vessel segmentation

Abstract Feature extraction of the retinal blood vessel is one of the crucial tasks in the prediction of ophthalmologic diseases. Important features are extracted based on image segmentation results. The efficiency of vessel segmentation methods could help doctors in the diagnostic of several relevant diseases as early as possible. Recently, U-Net has achieved good results in many medical image segmentation tasks, especially for images of blood vessels. However, due to the limitation of the network structure, some small features could be lost in the transmission process. As a result, there are still many research gaps for U-Net-based retinal vessel segmentation works. In this paper, we propose an improved U-Net based model to segment images of retinal vessels. The improvement focuses on U-Net from two aspects: designing a local feature enhancement module composed of dilated convolution and $$1\times 1$$ 1 × 1 convolution to enhance the feature extraction of tiny vessels; integrating an attention mechanism with skip connection of the network to highlight features related to vessel segmentation information taken from the down-sampling part to the up-sampling part. The performance of the proposed model was evaluated and compared with several published state-of-the-art approaches on the same public dataset—DRIVE, and the proposed method achieved an accuracy of 0.9563, F1-score of 0.823, TPR of 0.7983, and TNR of 0.9793. The AUC of PRC is 0.9109 and the AUC of ROC is 0.9794. The results proved the potential for clinical applications