<p>Skin cancer is among the most common and dangerous forms of cancer worldwide. The earlier stage lesions, if not diagnosed on time, transform into cancerous lesions. Therefore, early and accurate diagnosis is critical. These lesions are often misdiagnosed due to their varied shapes and sizes as well of poor contrast. In this study, we handled the problem by applying a CLAHE for contrast enhancement and then performed the data augmentation to handle the class imbalance. Afterwards, segmentation is performed using Attention U-Net and on the basis of segmented images, the contours are drawn on the boundary of lesions using Convex Hull and the Suzuki-Abe approach. To ensure an accurate highlighted boundary of lesions, the intersection of drawn boundaries is taken. Afterwards, the lesion classification is performed using various deep learning algorithms, including ResNet-18, ResNet-50, MobileNetV2, InceptionNet and XceptionNet. All the experiments were performed on the ISIC 2019 dataset, a publicly available benchmark dataset of dermoscopic images. The results showed that the given process helped improve the classification performance in all models. An inceptionNet architecture achieved the highest classification accuracy of 98.9%, with a precision of 98.5% and a recall of 98.4%, while other models, such as ResNet-50 and XceptionNet reached accuracy of 97.3% and 97.8% respectively. To validate and assess the model’s visual performance after the contours, Explainable AI is applied using LayerGrad-CAM and Grad-CAM++, making it easier to understand how the models processed the images on each layer. Overall, the proposed method improves both classification performance and the ability to explain the model’s decisions, that shows a potential for assisting in skin lesion classification.</p>