A ROBUST U-NET BASED FRAMEWORK FOR HIGH-FIDELITY SKIN LESION SEGMENTATION

Abstract

Automated segmentation of skin lesions remains a central challenge in dermatological imaging owing to substantial variability in lesion morphology, colour distribution and acquisition artefacts. This study presents a lightweight yet high-precision U-Net framework that integrates a hybrid Dice-binary cross-entropy optimisation scheme together with an expanded evaluation protocol incorporating boundary-aware metrics, offering improved robustness without architectural complexity. Trained and validated on 2594 dermoscopic images from the ISIC 2018 benchmark, the model achieved state-of-the-art performance for standard U-Net configurations, with a Dice coefficient of 0.94, intersection-over-union of 0.89, precision of 0.96, recall of 0.93, F1 score of 0.94, average symmetric surface distance of 5.4 pixels and 95th-percentile Hausdorff distance of 12.7 pixels. Qualitative inspection confirmed strong boundary localisation and stable behaviour across irregular, low-contrast and clinically challenging lesions. These findings demonstrate that substantial performance gains can be achieved through optimised training dynamics and a rigorously structured evaluation pipeline, establishing a reproducible and computationally efficient baseline that can support future methodological development in automated dermatological analysis.