Synthetic Melanoma Image Generation and Evaluation Using Generative Adversarial Networks.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computational Biology

Background:

• Melanoma detection relies on early diagnosis, with dermoscopy and deep learning showing promise.
• Limited datasets and class imbalance (few melanoma examples) hinder AI model development.
• Generative Adversarial Networks (GANs) offer potential for synthetic data generation.

Purpose of the Study:

• To systematically benchmark GAN architectures for high-resolution melanoma image synthesis.
• To evaluate image quality using quantitative metrics, qualitative inspection, and downstream task performance.
• To assess the utility of synthetic melanoma images in mitigating class imbalance for improved AI detection.

Main Methods:

• Compared four GANs (DCGAN, StyleGAN2, StyleGAN3-T, StyleGAN3-R) for 512x512 melanoma synthesis.
• Trained and optimized models on ISIC 2018 and ISIC 2020 datasets with unified preprocessing.
• Assessed image quality via FID, FMD, visual inspection, classification by a frozen EfficientNet, and dermatologist evaluation.

Main Results:

• StyleGAN2 demonstrated the best performance, balancing quantitative metrics and perceptual quality (FID: 24.8/7.96).
• A classifier identified 83% of StyleGAN2 images as melanoma; dermatologists achieved 66.5% accuracy in distinguishing real/synthetic images.
• Augmenting datasets with StyleGAN2-generated images improved melanoma detection AUC from 0.925 to 0.945.

Conclusions:

• StyleGAN2 effectively synthesizes diagnostically relevant melanoma images.
• Generated images can significantly improve AI model performance by addressing class imbalance.
• This approach offers a valuable tool for enhancing melanoma detection pipelines.