Skin cancer is among the most prevalent forms of cancer globally, and early detection plays a critical role in improving patient outcomes. This thesis presents a comprehensive deep learning-based framework for the automated classification of dermoscopic skin lesion images using the HAM10000 dataset. The proposed methodology involves multiple stages, including image preprocessing, data segmentation, and the application of advanced convolutional neural network (CNN) architectures through transfer learning models. The performance of seven cutting edge pretrained models—ResNet50, MobileNetV2, VGG19, DenseNet121, InceptionV3, Xception, and EfficientNetB0—was assessed and refined separately. Weighted loss functions and class-balancing techniques were used to address the dataset’s intrinsic class imbalance problem. Additionally, a performance-based weighted soft voting mechanism was used to integrate all base learners to create a unique Uncertainty-Guided Ensemble Model. To increase prediction accuracy and model calibration, uncertainty quantification methods including temperature scaling and confidence thresholding were used. Compared to individual models, the ensemble model performed better, outperforming the best standalone model (EfficientNetB0) by 1.68% in accuracy and 1.95% in F1 score, with an accuracy of 85.42% and an F1 score of 85.93%. A wide range of metrics, including accuracy, recall, specificity, F1-score, and AUC-ROC, were used for evaluation, both overall and for each class. This study opens the door for practical implementation in computer aided dermatological diagnosis systems by proving the effectiveness of integrating ensemble learning, transfer learning, and uncertainty modelling for accurate and dependable skin lesion classification.# Skin_lesion_classification