GAN-Based Data Augmentation and Anonymization for Skin-Lesion Analysis: A Critical Review

TL;DR

This paper critically reviews the use of GANs for data augmentation and anonymization in skin-lesion analysis, highlighting their limited effectiveness mainly on out-of-distribution test sets and cautioning their medical application.

Contribution

It provides a comprehensive evaluation of GAN-based methods in skin-lesion analysis, emphasizing the need for careful consideration of their limitations and risks.

Findings

GANs improve performance mainly on out-of-distribution test sets

GAN-based anonymization shows limited benefits for in-distribution data

Caution is advised due to costs and potential risks of GAN deployment in medical settings

Abstract

Despite the growing availability of high-quality public datasets, the lack of training samples is still one of the main challenges of deep-learning for skin lesion analysis. Generative Adversarial Networks (GANs) appear as an enticing alternative to alleviate the issue, by synthesizing samples indistinguishable from real images, with a plethora of works employing them for medical applications. Nevertheless, carefully designed experiments for skin-lesion diagnosis with GAN-based data augmentation show favorable results only on out-of-distribution test sets. For GAN-based data anonymization $-$ where the synthetic images replace the real ones $-$ favorable results also only appear for out-of-distribution test sets. Because of the costs and risks associated with GAN usage, those results suggest caution in their adoption for medical applications.