This chapter presents a new method based on the U-Net architecture for the automatic segmentation of the major temporal arcade (MTA) in retinal fundus images. The method consists of the steps of segmentation and attention. In the first step, the U-Net architecture is used to segment vessel-like structures vía transfer learning with the well-known VGG-16 deep learning model. Since the MTA is the thickest blood vessel structure in the retina and it can be considered as a parabolic shape, in the second step, an attention module is introduced in order to capture the MTA shape removing all the vessel-like structures different to the MTA. The performance of the proposed method is evaluated in terms of sensitivity, specificity, accuracy and F1-score, and it is compared with different state-of-the-art vessel segmentation methods. The database of retinal fundus images is the DRIVE database, and the MTA ground-truth images were labeled by a specialist in order to isolate that blood vessel structure. The database of MTA images was divided into training and testing sets, where a data augmentation strategy of rotation was only implemented on the training set in order to increase the number of samples of the proposed and comparative methods. In the experiments, the numerical results obtained by the proposed method in the automatic MTA segmentation problem outperforms the comparative methods obtaining an accuracy of 0.9889 and F1-score of 0.6413 using the testing set of 18 retinal fundus images. In addition, the average computational time of the proposed deep learning U-Net model in testing images is 0.003 s using specialized hardware .