Melanoma and other skin cancers can be detected early by recognizing suspicious skin lesions. Although traditional U-Net models are widely utilized in medical picture segmentation because of their capacity to catch tiny details, they frequently struggle to comprehend larger contextual information and effectively delineate complex lesion boundaries. In this research to tackle these issues, we introduced a hybrid approach that improves the accuracy and efficacy of skin lesion identification by using Attention U-Net for segmentation and ResNet-50 for feature extraction. While Attention U-Net’s attention mechanism makes sure the model concentrates on the most crucial regions for segmentation, ResNet-50 employs this method to extract significant features from the image. Residual connections also serve to stabilize training by preventing vanishing gradient concerns, which makes the network more dependable and successful at recognizing skin lesions. We tested our model on two popular data sets, ISIC 2019 and PH2, using pre-processing strategies such as Black-Hat filtering to eliminate the hair and data augmentation to enhance the model’s generalizability. In order to evaluate its efficiency, we used two crucial metrics are Dice Similarity Coefficient (DSC) and the Jaccard Index.