Skin cancer, a prevalent and potentially deadly condition resulting from factors such as UV rays exposure and ozone layer depletion, is a significant health concern worldwide, with approximately 5 million cases reported annually. Its gravity lies in its ability to metastasize to the brain, causing complications like headaches, seizures, and paralysis. Early detection plays a pivotal role in mitigating mortality rates, yet subjective clinical assessments for lesion segmentation often fall short of the expectations of expert dermatologists. Biomedical image segmentation, which involves the automatic or semi-automatic delineation of boundaries in 2D images, encounters challenges due to variations in shape, location, size, and texture within medical images. Manual segmentation, a laborious and time-consuming endeavor, underscores the need for fully automated approaches. In this context, this paper introduces a UNet-based architecture enhanced by the pre-trained ResNet50 model. The comprehensive results showcase outstanding performance across five evaluation parameters for medical image segmentation. Notably, this GA-UNet approach surpasses traditional methods with an impressive accuracy rate of 98.7% and a Dice Similarity Coefficient of 89.6%.