Skin cancer is one of the most epidemic cancers worldwide, which needs early and accurate detection. Although issues such as high computational costs, lack of interpretability, and class imbalance remain, recent developments in deep learning have opened up new possibilities for medical image classification. By combining MobileNetV3, EfficientNetB0, and a customized convolutional neural network (CNN), we propose a lightweight CNN architecture for the categorization of skin lesions. Our goal is to obtain high accuracy while minimizing computational complexity so that our model can be deployed in edge devices or resource-constrained environments. Starting with pixel values stored in a CSV file in the HAM10K dataset, we reconstructed the images and then resized them into three different input resolutions. After combining attention modules such as Selective Kernel (SK), Convolutional Block Attention Module (CBAM), and Squeeze-and-Excitation (SE) with lightweight base models to improve feature extraction, we got 97.66%, 98.44%, 98.57% individually for MobileNetV3, EfficientNetB0, and custom CNN respectively. We beat individual models with a remarkable classification accuracy of 99.47% using our proposed ensemble technique. We used explainable artificial intelligence (XAI) methods such as Grad-CAM and LIME to improve the transparency, trust, and confidence of the model. These techniques ensured attention to clinically significant regions by offering visual insights into the model’s predictions. This work shows how lightweight CNNs with attention and XAI approaches can effectively detect skin cancer early, providing a transparent and extensible approach for efficient applications in the healthcare industry.