Improved Segmentation of Skin Lesions Using Attention-Enhanced Residual U-Net

Accurate segmentation of skin lesions is critical for early skin cancer identification and treatment. Conventional U-Net architectures have been widely utilized in medical image segmentation for their ability to capture intricate details effectively. However, these models often encounter challenges in grasping contextual information and managing complex lesion boundaries. We propose an enhanced segmentation model, the attention-based residual U-Net, which incorporates attention mechanisms and residual connections into the traditional U-Net architecture. The attention mechanism enables the network to concentrate on relevant regions of the input image, improving feature extraction, while residual connections aid in training deeper networks by tackling the issue of vanishing gradients. Assessed on a publicly available dermoscopic image dataset (PH2 dataset), our model demonstrates significant improvements in segmentation accuracy (96.55%) and boundary outlining, attaining higher dice coefficients (89.54%) and reduced segmentation errors. The proposed model displays resilience across various lesion variations and imaging conditions, making it a promising tool for clinical applications in dermatology.