Accurate and interpretable skin lesion classification is crucial for effective skin cancer diagnosis. However, conventional convolutional neural networks (CNNs) and attention models often lack interpretability and tend to rely on misleading correlations. We propose a Gradient-Guided Causal Attention (GGCA) framework that advances skin cancer classification by integrating a ResNet-50 backbone with a gradient-based attention mechanism. This mechanism dynamically generates spatial masks using classification loss gradients, focusing on causally significant lesion regions to improve robustness and transparency. Our approach integrates causal Grad-CAM visualizations to provide clinically relevant explanations aligned with expert annotations, enabling precise diagnosis across lesion types. Evaluated on the ISIC 2019 (eight classes) and HAM10000 (seven classes) datasets, our method achieves state-of-the-art (SOTA) performance. Specifically, it obtains 0.939 precision, 0.937 recall, 0.935 F1-score, and 0.939 accuracy on ISIC 2019, and 0.917 precision, 0.915 recall, 0.915 F1-score, and 0.917 accuracy on HAM10000, surpassing the baselines by 7–9% across all evaluation metrics. This framework establishes a new benchmark for reliable and interpretable skin lesion diagnosis and advances trustworthy AI for clinical adoption.