Abstract
Automatic skin lesion segmentation is one of the key pivotal tasks in dermatological image processing, with important consequences in early melanoma diagnosis and treatment planning. Nevertheless, issues of extreme class imbalance, morphological variability, and poor boundary delineation still exist in current deep learning-based approaches. This study proposes a unified framework that combines morphology-guided latent interpolation and synthesis for lesion augmentation (MoG-LISA) and CB-SwinGMO (Class-Balanced Swin-UNet optimization using geometric mean-driven feedback evolutionary framework) to address these challenges. MoG-LISA generates high-fidelity synthetic samples in a morphology-aware latent space, enriching underrepresented lesion classes such as melanoma and vascular anomalies. Meanwhile, CB-SwinGMO employs multi-objective evolutionary optimization to adapt Swin-UNet parameters for improved generalization and precise boundary detection. Quantitative results highlight the superior performance of our approach, achieving a dice similarity coefficient (DSC) of 93.8%, IoU of 91.2%, boundary accuracy of 92.7%, and Hausdorff distance reduction up to 4.3 pixels on the SIIM-ISIC dataset.