Accurate multi-class classification of skin lesions remains a challenging task due to the intrinsic complexity of dermoscopic images and the severe class imbalance present in medical datasets. This paper introduces an ensemble-based deep learning pipeline designed to achieve robust and generalizable multi-class skin lesion diagnosis. The proposed framework systematically addresses the primary challenges of class imbalance using an aggressive mix up augmentation. A five face structured pipeline encompassing data preparation, stratified crossvalidation, model training, ensemble inference, and performance evaluation is performed. The proposed ensemble integrates five heterogeneous deep architectures: Vision Transformer (ViT), Swin Transformer, ConvNeXt, EfficientNet, and DenseNet, each contributing complementary representational strengths. To enhance model robustness and fairness, the pipeline incorporates regularization and a hybrid loss combining focal Loss and label smoothing. Ensemble predictions are aggregated using a weighted soft voting strategy, ensuring stable and accurate classification outcomes across all lesion categories. Experimental evaluations demonstrate that the proposed system achieves state-of-the-art performance on the HAM10000 dataset, delivering high accuracy and balanced sensitivity across minority classes. The end-to-end design emphasizes reproducibility, scalability, and transparency, establishing a robust foundation for future research in automated dermatological diagnosis.

Modern neural networks have recently been found to be poorly calibrated, primarily in the direction of over-confidence. Methods like entropy penalty and temperature smoothing improve calibration by clamping confidence, but in doing so compromise the many legitimately confident predictions. We propose a more principled fix that minimizes an explicit calibration error during training. We present MMCE, a RKHS kernel based measure of calibration that is efficiently trainable alongside the negative likelihood loss without careful hyper-parameter tuning. Theoretically too, MMCE is a sound measure of calibration that is minimized at perfect calibration, and whose finite sample estimates are consistent and enjoy fast convergence rates. Extensive experiments on several network architectures demonstrate that MMCE is a fast, stable, and accurate method to minimize calibration error while maximally preserving the number of high confidence predictions.