Res2Unet: A multi-scale channel attention network for retinal vessel segmentation

Retinal diseases can be found timely by observing retinal fundus images. So extracting blood vessels from retinal images is an important part because it is the way to show the changes of vessels. However, most of the previous methods based on deep learning cared more about accuracy and ignored the complexity of the model for segmenting retinal vessels, which makes these methods difficult to apply to medical equipment. Besides, due to the great differences in the width of retinal vessels, some methods cannot well-extract all blood vessels at the same time. Based on above limitations, we propose a new lightweight network, called Res2Unet. It applies a multi-scale strategy to extract blood vessels of different widths and integrates the strategy into the channels to greatly reduce parameters and computation resources. Res2Unet also uses channel-attention mechanism to promote the communication between channels and recalibrate the relationship of channel features. Then, we propose two post-processing methods. One called the local threshold method(LTM) uses a lower local threshold to excavate hidden blood vessels in discontinuous blood vessels of the probability maps. The other named weighted correction method (WCM) combines the probability maps of Unet and Res2Unet to remove false positive and false negative samples. On the DRIVE dataset, the Dice, IOU and AUC of our Res2Unet reach 0.8186, 0.6926 and 0.9772, respectively, which are better than that of Unet with 0.8109, 0.6817 and 0.9751. Importantly, the number of parameters of Res2Unet are about one-third of Unet. It means that Res2Unet has less hardware requirements.