Evaluating the Utility of Synthetic Image Generation for Medical AI: A Review

Abstract

In recent years, synthetic data generation techniques have shown great potential in generating realistic data. In healthcare, synthetic data can help address many challenges, including privacy concerns, improving accessibility of training datasets, and reducing bias. This study explores different synthetic image generation techniques by reviewing high-quality peer-reviewed articles. These articles are chosen based on our devised inclusion and exclusion criteria. The generation techniques mainly consist of Generative Adversarial Network (GAN) and its variants, followed by Variational Autoencoders (VAEs), diffusion models and 3D simulation. The study findings show that synthetic data can enhance the performance of AI models by improving the quality of existing datasets. However, there are still some limitations, such as unstable training, mode collapse, lack of effective evaluation metrics and explainability and high computational cost, that need to be addressed to unlock the full potential of generative models.

Ähnliche Arbeiten