Optimizing Skin Cancer Diagnosis: DenseNet-121 for Enhanced Binary Classification

Purpose Timely diagnosis and treatment depend on the early and precise classification of skin lesions as benign or malignant. This research attempts to create a binary-classification system based on deep learning that aligns with real-world clinical decision-making, using a curated dataset to enhance model relevance and interpretability. Methods A customized version of the PAD-UFES-20 dataset comprising 2298 dermoscopic images was restructured from six diagnostic categories into two classes—benign and malignant. The DenseNet-121 convolutional neural network was employed as the core architecture, equipped with a custom classification head. Preprocessing included image resizing, normalization, and real-time data augmentation. Results The proposed DenseNet-121 model achieved a validation accuracy of 95.33%, outperforming the benchmark EfficientNetB4 model, which achieved 93.00% under the same conditions. The model also demonstrated strong generalization capability, with low false negative rates observed through confusion matrix analysis. Conclusion The combination of dataset restructuring, architectural optimization, and regularization strategies resulted in a high-performing, efficient skin cancer classification model. The system shows promise for deployment in clinical and teledermatology settings, with future work focused on explainability and deployment optimization.