Skin lesion segmentation is a critical step in the early detection of skin cancer, a condition with rapidly increasing global incidence. This study introduces an advanced method leveraging the complementary strengths of Multi-resolution UNet and K-Fold cross-validation within a deep learning framework. The multi-resolution architecture excels in capturing intricate lesion patterns through enhanced feature extraction, while K-Fold cross-validation ensures robust training and minimizes overfitting, improving generalizability across datasets. Using the PH2 dataset of dermoscopic images, the proposed approach demonstrates significant improvements in segmentation accuracy compared to traditional UNet and state-of-the-art methods. The inclusion of K-Fold validation not only ensures reliable performance metrics but also reinforces the model’s robustness. Experimental results highlight the superior performance of the Multi-resolution UNet, emphasizing its potential for clinical dermatology applications and underscoring the value of innovative architectures and rigorous validation in medical image analysis. The model shows strong performance with accuracy (0.9733) and precision (0.9738) indicating few errors. IoU (0.7287) and Dice (0.6101) show good but moderate overlap, while recall (0.9434) reflects high detection. Overall, the model performs well with slight room for improvement.