Abstract: Melanoma is one of the highly aggressive types of skin cancer, and early detection plays a critical role in patient survival. Although dermoscopy exposes subtle lesion patterns for clinical evaluation, manual diagnosis can be slow and heavily dependent on specialist expertise. Deep-learning methods have improved automated melanoma identification substantially, yet many current models still encounter real-world challenges, including noisy inputs, non-lesion images, variable lighting conditions, and class imbalance. To address these limitations, this study introduces Melanoma Spotter, a dual-stage diagnostic system that combines a lesion-validation network with a hybrid VGG16–DenseNet121 classifier. The validation stage removes irrelevant or low-quality images to ensure that only true dermoscopic data proceeds to analysis, while the fused CNN model exploits the complementary strengths of both architectures to produce more stable and accurate predictions. Experimental results on dermoscopic datasets show improved reliability, higher confidence calibration, and better robustness than standalone networks. Overall, Melanoma Spotter demonstrates strong promise as a practical deep-learning solution to support early detection of melanoma in clinical settings.

Keywords: VGG16, DenseNet121, HAM10000, Deep Learning, Melanoma, Hybrid Model, Skin Lesion Detection

Downloads: | DOI: 10.17148/IARJSET.2026.13110