Precise skin lesion localization of dermoscopic images is vital in the early detection of melanoma and digital dermatology processes. This paper is a full assessment of three light and interpretable deep learning architectures, namely U-Net, a modified convolutional neural network (CNN), and a transformer-based SegFormer, on the HAM10000-ISIC2018 dataset. The preprocessing, enhancement, and augmentation conditions were the same in all the models to succeed the fairness and repeatability. It included Grad-CAM and attention-rollout explainability methods to visualize areas of focus in models and choose to make them more clinical interpretable. Dice score, precision, and validation accuracy quantitatively show that both U-Net and SegFormer attained the highest Dice score 93.05%, and the custom CNN, next in terms of architectural complexity, has attained a competitive Dice score of 88.22%. SegFormer was the most precise 95.51% and effective, which means that it is very suitable in resource-limited or mobile diagnostic settings. [NEW] We validated model robustness against common dermatological artifacts (hair, ruler marks, gel bubbles) and compared performance with recent state-of-the-art architectures including TransUNet and Swin-UNet. The experimental results show that light and understandable segmentation models can be used to reach clinically significant performance and can be used as a reliable element in future computer-aided dermatology systems and diagnostic piping.