Skin lesions encompass alterations or irregularities of the skin, including moles, cysts, freckles, and similar conditions. While many skin lesions are benign, certain types can signal or progress to skin cancer, a disease whose incidence has risen significantly in recent years, leading to many deaths. Although dermatologists can identify skin cancers by manual examination, this process is often costly and time-consuming. Several research projects increasingly aim to create automated systems for early skin cancer detection, to help reduce its mortality rate. This study introduces a deep-learning approach utilizing convolutional neural networks to classify skin lesions. It aims to categorize lesions into seven classes using the challenging dataset HAM10000. We adopt a new iterative thresholding algorithm-based optimizer to optimize the proposed model. The latter dataset is imbalanced, to reduce this issue and to enhance the model performance, we proceed to offline data augmentation by using different image transformations. After fine-tuning the model’s hyper-parameters, we achieved accurate classification. The test results demonstrate an accuracy of 92.62% and macro-average ROC-AUC score of 98.55%. Given these encouraging results, this model can help doctors diagnose skin lesions while saving processing time.