Skin cancer poses a significant public health challenge, where computer-based automated systems can play an important role in easing the growing burden of this disease. Accurate segmentation of skin lesions from dermoscopic images is a vital step toward effective diagnosis. However, this task remains challenging due to variations in lesion shape and contrast, the presence of natural and artificial artifacts such as hair and air bubbles, image noise, and inconsistencies in image acquisition settings. This study presents an Attention Residual U-Net model for semantic segmentation of skin lesions in dermoscopic images. The architecture employs skip connections to preserve spatial details lost during encoding. The integration of attention mechanisms and residual skip connections enables the network to capture fine-grained contextual and spatial details, thereby enhancing lesion boundary detection and segmentation. Evaluation metrics such as Jaccard Coefficient, Dice Coefficient and its loss function are used to assess the models’ performance. Additionally, the study examines the influence of various hyperparameters, noise factors, and preprocessing strategies on performance. Experimental results on the ISIC (International Skin Imaging Collaboration) 2018 dataset demonstrate that the proposed model surpasses existing approaches. The model achieved a segmentation accuracy of 94.8%, and maintained strong robustness under Gaussian and salt-and-pepper noise with accuracies of 89.9% and 91.1%, respectively.