Deep Learning Architecture Optimization for Skin Cancer Image Classification on Multi-Source Dataset

Early detection of skin cancer is crucial to reduce diagnostic delays and improve patient outcomes, yet existing automated systems often suffer from limited generalization due to single-source datasets and restricted model evaluation. This study presents a deep learning based skin cancer image classification system using a multi-source dataset of 13,902 dermatoscopic images categorized into benign and malignant classes. In the first phase, a benchmarking study was conducted on seven pre-trained CNN architectures, namely MobileNetV2, InceptionV3, Xception, DenseNet169, ResNet50, VGG16, and VGG19. The results indicate that DenseNet169 achieved the best baseline performance with a test accuracy of 89.30%. In the second phase, the DenseNet169 architecture was optimized through structural modification of dense layers, application of dropout regularization, and selective fine-tuning of backbone layers. The optimized model improved the test accuracy to 91.20% and achieved an AUC-ROC of 97.14%, demonstrating enhanced robustness and sensitivity in detecting malignant lesions. The novelty of this work lies in the integration of multi-source datasets combined with targeted architectural optimization of DenseNet169 to reduce false-negative rates in malignant detection. These findings highlight the potential of the proposed model as a reliable non-invasive clinical decision support tool for early skin cancer diagnosis, and emphasize the need for further validation using prospective clinical datasets and real-world deployment scenarios to ensure its practical applicability in clinical environments.