The accurate segmentation of retinal vessels is of paramount importance for the diagnosis of ophthalmic diseases and the detection of diabetes at its earliest stages. Despite the remarkable efficacy of numerous deep learning methodologies, segmentation errors remain, particularly in the segmentation of vessel terminals. This study proposed MRFFA-Net, which is an innovative network model integrating attention mechanisms and multi-scale residuals to address the aforementioned challenge. A novel multi-scale residuals attention block is proposed, which combines global and local information to enhance the precision of vessel terminal segmentation. Furthermore, a ShuffleAttention block was incorporated to learn the structures of both vessels and non-vessels, thereby enhancing the model’s generalization capability. To validate the effectiveness of our method, experiments were conducted on the DRIVE dataset, achieving an AUC of 98.04% and an F1-score of 82.6%.