<p>Skin cancer is a potentially fatal disease that requires early and accurate diagnosis to improve patient outcomes. Deep learning has shown promise in automating skin lesion classification; however, many existing models suffer from limited interpretability, class imbalance, and irrelevant feature extraction. This study introduces CBAM-Xception, an explainable attention-guided deep learning model designed to enhance classification performance by focusing on clinically relevant lesion features. The model integrates a pretrained Xception backbone with a Convolutional Block Attention Module (CBAM) to highlight discriminative regions while suppressing background noise. CLAHE was applied to enhance contrast in dermoscopic images, and geometric and color augmentation were used to address class imbalance in the HAM10000 (seven classes) and ISIC 2019 (nine classes) datasets. The evaluation was performed using a standard training, validation, and test split. The first 50 layers of Xception were frozen before fine-tuning. Grad-CAM++ visualizations confirmed the focus of the model on key lesion areas. The proposed model achieved superior performance compared to the MobileNet and EfficientNet baselines, attaining accuracies of 98.62% (AUC: 0.9997) on HAM10000 and 93.66% (AUC: 0.9939) on ISIC 2019, demonstrating its strong effectiveness. However, the model’s performance may depend on dataset-specific characteristics and computational resources, which could limit its generalizability to unseen clinical environments. By combining high accuracy, interpretability, and robustness to class imbalances, CBAM-Xception provides a reliable solution for automated skin cancer diagnosis.</p>