Skin cancer is one of the fatal illnesses that cause a high percentage of deaths worldwide. The treatment of skin cancer depends critically on early detection of the disease. Dermatologists face several difficulties in the diagnostic process to address this issue. This study aims to improve the classification accuracy of seven types of skin tumors by using transfer learning on pre-trained Convolutional Neural Network models. The modified models, including EfficientNetB3, EfficientNetB0, and ResNet50, were trained on a dataset of 10,015 dermatoscopic images. The dataset was augmented using data augmentation techniques, and the class imbalance problem was addressed. This study offers a promising approach for clinicians to make informed judgments about patient diagnoses. Our results showed the superiority of the modified models over the official models. Our best model, the modified EfficientNetB3, achieved an accuracy of 91.6%, Precision of 83%, F1-Score of 88 %, and Recall of 94% on the HAM10000 dataset.