Improved U-Net Fundus Image Segmentation Algorithm Integrating Effective Channel Attention

Abstract

Fundus blood vessel segmentation is important to obtain the early diagnosis of ophthalmic- related diseases. A great number of approaches have been published, yet micro-vessel segmentation is still not able to deliver the desired results. In this paper, an improved retinal segmentation algorithm incorporating an effective channel attention (ECA) module is presented. Firstly, the ECA module is imported into the downsampling stage of a U-shape neural network (U-Net) to capture the cross-channel interaction information. Secondly, a dilated convolutional module is added to expand the receptive field of the retina, so that more micro-vessel features can be extracted. Experiments were performed on two publicly available datasets, namely DRIVE and CHASE_DB1. Finally, the improved U-Net was used to validate the results. The proposed method achieves high accuracy in terms of the dice coefficient, mean pixel accuracy (mPA) metric and the mean intersection over union (mIoU) metric. The advantages of the algorithm include low complexity and having to use fewer parameters.