CNN has demonstrated exceptional effectiveness in deep learning applications, especially in medical image classification. This research investigates how various CNN architectures perform in melanoma classification using dermoscopic images, highlighting the efficacy of pre-trained models such as ResNet50, VGG16, InceptionV3, and MobileNetV2. A balanced dataset consisting of 3300 dermoscopic images from the ISIC repository was used for training, containing an almost equal distribution of 1675 benign and 1625 malignant cases. To improve the effectiveness of the model and avoid overfitting, different data augmentation were applied. A thorough assessment was conducted for each model, enabling the precise determination of key evaluation metrics, such as classification accuracy, precision, recall, and F1-score, effectively demonstrating their performance. MobileNetV2 showed exceptional performance, reaching an accuracy of 90.14%, a strong candidate for melanoma identification. Additionally, the study employed the confusion matrix and the ROC curve to comprehensively assess the classification of models. MobileNetV2 demonstrated strong discriminatory power, achieving an AUC of 0.96, proving its usefulness in classifying between malignant and benign cases. This comparative analysis highlights MobileNetV2’s effectiveness in melanoma classification and the advancements convolutional neural networks bring to diagnostic methods.