Skin lesion classification plays a crucial role in dermatology, facilitating the early detection and treatment of malignant conditions such as melanoma, which significantly improves treatment outcomes and reduces mortality risk. This study introduces two models built on advanced pre-trained convolutional neural networks, ResNet50 and Inception-ResNetV2. The proposed models are evaluated on the benchmark HAM10000 dataset with the skin images undergoing preprocessing and augmentation to enhance classification accuracy. Both models are implemented with two configurations: (A) full-model training, where all layers are fine-tuned to adapt to the task-specific data, and (B) training only the newly added classification layers while keeping the pre-trained layers frozen. Various optimization and regularization techniques are integrated to improve model performance. The experimental results demonstrate that the models trained under configuration A outperform those under configuration B and achieve competitive and promising results compared to the state-of-the-art approaches in skin lesion classification. Further analysis, including learning curves, confusion matrices, and hyperparameters, is conducted to validate the efficiency and robustness of the proposed models.