Skin cancer diagnosis is a critical task where the accuracy of automated tools can directly impact patient outcomes. This study proposes the comparison of five popular pre-trained convolutional neural network (CNN) architectures including VGG19, ResNet50, Efficient-NetB0, MobileNetV2, and DenseNet121 to determine the best performing model in classifying skin lesions as benign or malignant. We curated a well-balanced dataset by merging and downsampling the ISIC 2018, 2019, and 2020 datasets and used a standard preprocessing and augmentation pipeline to provide an unbiased comparison and enhance model generalization. All models were trained on resized and normalized images, with extensive augmentations to simulate a more heterogeneous dataset. The extensive experimental results demonstrated among the models DenseNet121 was the most effective, with the highest accuracy 90.3% and F1-score for malignant cases and possessing high generalization capability. In addition, MobileNetV2 performed well in terms of speed and efficiency, although with a lower recall.