Comprehensive Summary
This study investigates the use of advanced generative adversarial networks (GANs) to identify and classify suspicious naevi, mole-like skin lesions with melanoma risk. Using 33,000 dermoscopic images from 59 patients, the researchers worked to categorize lesions as suspicious or non-suspicious. Using their data, they developed three deep learning models: Deep Convolutional GAN (DCGAN), Auxiliary Classifier GAN (ACGAN), and a novel Variational Autoencoder ACGAN (VAE-ACGAN). While all models produced realistic synthetic skin lesion images, VAE-ACGAN achieved the highest performance, clearly separating suspicious and non-suspicious naevi. Statistical testing confirmed that this model outperformed ACGAN in accuracy, sensitivity, and AUC values. Additionally, the model revealed clinical variability between clustering non-suspicious naevi and more dispersed suspicious naevi, which could also have clinical implications
Outcomes and Implications
Early detection of melanoma greatly improves survival and suspicious naevi are key indicators of risk. Since current manual diagnosis by professionals is resource-intensive and subjective, there is a need for novel, reliable automated systems. By generating high-quality training data and providing interpretable lesion clusters, VAE-ACGAN could enhance clinical decision making and improve diagnostic tools. While research is still being done on these models, their findings suggest that similar deep learning models may soon support clinical workflows by expanding datasets, reducing observer variability, and improving early melanoma screening accuracy.