Abstract
In this study, six convolutional neural network (CNN) architectures, VGG16, Inception-v3, ResNet, MobileNet, NasNet, and EfficientNet are tested on classifying dermatological lesions. The research preprocesses and features extracts skin lesions data to achieve an accurate skin lesion classification in employing two benchmark datasets, HAM10000 and ISIC-2019. The CNN models then extract features from the filtered, resized images (uniform dimensions: 128 × 128 × 3 pixels). These results show that EfficientNet consistently achieves higher accuracy, precision, recall, and F1-score than any other model on melanoma, basal cell carcinoma and actinic keratoses, with 94.0%, 92.0%, 93.8%, respectively. The competitive performance of NasNet is also demonstrated for eczema and psoriasis. This study concludes that proper preprocessing and optimized CNN architecture are important for dermatological image classification. The results are promising, however, challenges such as the imbalance in the datasets and the requirement for larger ethically gathered datasets exist. For future work, dataset diversity will be improved, along with model generalization, through interdisciplinary collaboration and advanced CNN architectures.