Early diagnosis of skin cancer, especially melanoma, is important in saving lives of patients. Despite the fact that deep learning has enhanced automated lesion classification, the current models are characterized by high variance, low domain robustness, and low interpretability. We introduce a explainable ensemble which is a dense net 121 plus inception v3 plus a dense net plus squeeze-and-excitation (SE) attention to learn both complementary channel and spatial features. Lesion-specific regions are visualized in Grad-CAM with the guidance of the predictions, which boosts clinical trust. The experiments on the HAM10000 dataset demonstrate that the ensemble has 97.41% accuracy, which is higher than the individual models and state-of-the-art methods. This framework offers a credible and clear solution to Computer-Aided Diagnosis (CAD) in dermatology.