Integrating Machine Learning and Deep Learning for Improved Melanoma Diagnosis Using a Robust Transfer Learning Framework

The most common cause of skin cancer-related mortality worldwide is melanoma, the most severe kind of the disease. Patients' chances of survival are significantly increased when skin cancer is discovered early. Manual dermoscopic examination, however, is arbitrary and depends on the judgment of a dermatologist. Convolutional neural networks (CNNs), in particular, are deep learning models that have demonstrated great potential for automatically detecting melanoma. However, a number of studies show shortcomings in clinically meaningful evaluation criteria, robust repeatability mechanisms, and comparative architectural analysis. Using the HAM10000 dataset, this study offers a thorough transfer learning method for binary melanoma detection. Using a regulated and reproducible process, we evaluate three pre-trained CNN architectures: ResNet50, DenseNet121, and MobileNetV2. We provide mathematical formulae for evaluation metrics, transfer learning optimization, and convolutional procedures. Experimental results show that ResNet50 performs better discriminatively in ROC space, whereas DenseNet121 achieves the best sensitivity-specificity balance. The proposed architecture may be adapted for applications with several classes and fine-tuning, is repeatable, and is consistent with clinical practice.