Enhancing Skin Lesion Segmentation With AI: An Enhanced Scaled Dot Attention Mechanism Integrated With UNet for Melanoma Diagnosis

ABSTRACT

Accurate skin lesion segmentation is essential for computer‐aided melanoma diagnosis, as precise lesion boundary delineation directly influences clinical assessment and treatment planning. However, irregular lesion morphology, low‐contrast boundaries, varying lesion sizes and imaging artefacts remain significant challenges for existing deep learning methods. This paper proposes E‐SDAM‐UNet, an enhanced attention‐guided segmentation framework that extends the conventional UNet architecture by integrating four complementary modules: hierarchical multi‐scale feature fusion (HMSF) for contextual feature aggregation, an enhanced scaled dot attention mechanism (E‐SDAM) for adaptive spatial feature learning, efficient channel attention (ECA) for channel‐wise feature refinement and an edge‐aware boundary refinement module (EBRM) for improved lesion boundary localisation. A hybrid dice and binary cross‐entropy loss function is employed to address class imbalance and enhance optimisation. The proposed framework is evaluated on the HAM10000, ISIC 2018 and PH2 datasets, including cross‐dataset validation to assess generalisation. Performance is measured using dice, IoU, precision, recall, F1‐score, sensitivity, specificity, MCC, HD95 and ASD. Experimental results demonstrate that E‐SDAM‐UNet consistently outperforms representative state‐of‐the‐art segmentation methods while maintaining computational efficiency and robust cross‐dataset performance, making it a promising solution for automated dermoscopic image segmentation and computer‐aided melanoma diagnosis.