This paper introduces a novel approach for the automatic segmentation of the major temporal arcade (MTA) in retinal fundus images, which is based on the U-Net architecture. The segmentation and attention processes make up the approach. The first step involves segmenting vessel-like structures using the well-known VGG-16 deep learning model and transfer learning using the U-Net architecture. In the second stage, an attention module is introduced to capture the MTA form, eliminating all vessel-like structures that are not the MTA. The effectiveness of the proposed approach is assessed in terms of F1-score, sensitivity, specificity, and accuracy, and it is contrasted with various cutting-edge vascular segmentation techniques. Only the training set was subjected to a data augmentation strategy. To identify the blood vessel structure, a specialist classified the MTA ground-truth images from the well-known DRIVE database containing 40 retinal fundus images. To ensure a robust training process, data augmentation was applied, and each training image was divided into 81 patches, resulting in a total of 1,377 patches for training. 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 0.9923 and F1-score of 0.7700 using the testing set of retinal fundus images. In addition, the average computational time of the proposed deep learning model in testing images is 0.003 s using specialized hardware.