JMIR Publications

ABSTRACT

Melanoma is an aggressive form of skin cancer with rising incidence worldwide, making early detection critical for patient survival. This study proposes an automated deep learning framework for melanoma classification using a conventional Convolutional Neural Network (CNN) and two Dense Convolutional Network models DenseNet169 and DenseNet201. Publicly available dermoscopic images from Kaggle, representing multiple malignant and benign skin conditions, were preprocessed, resized to 250×250 pixels, and augmented to address class imbalance. Models were trained with Adam and RMSprop optimizers across 20, 60, and 100 epochs, and evaluated using accuracy, F1score, sensitivity, and specificity. Results show that DenseNet169 and DenseNet201 outperformed the baseline CNN, achieving up to 95% accuracy and higher F1scores after data augmentation, with reduced overfitting and improved generalization. These findings highlight the potential of DenseNet architecture, combined with targeted preprocessing and optimization strategies, to deliver accurate, efficient, and scalable melanoma detection systems. The proposed framework could be integrated into clinical workflows and teledermatology platforms, aiding physicians in rapid, reliable lesion assessment and supporting earlier intervention.