Comparative Analysis of Synthetic and Real Melanoma Images in AI-Driven Diagnosis

Abstract

A Computer-Aided Diagnosis system is a technology used in medical imaging to support clinicians in the interpretation of medical images, enhancing both the accuracy and efficiency of medical diagnoses through the integration of advanced AI technologies. In this paper, the approach involved training and testing four neural networks (AlexNet, ResNet, EfficientNetB0 and ViT) from scratch and with fine-tuning on the Refined ISIC dataset (Scenario I), and training on synthetic data and testing on Refined ISIC (Scenario II) for the Melanoma Binary Detection. Results showed that AlexNet, when trained from scratch, achieved the best performance in Scenario I with accuracy, sensitivity, specificity, precision, F1-SCORE, and false negative rate values of 82.2%, 75.8%, 87.4%, 83.1%, 79.3%, and 20.7%, respectively. For Scenario II, AlexNet with fine-tuning reached the highest performance with 73.4% accuracy, 62.8% sensitivity, 82.2% specificity, 74.5% precision, 68.2% F1-score, and a false negative rate of 37.2%.