The Beauty or the Beast: Which Aspect of Synthetic Medical Images Deserves Our Focus?

Training medical AI algorithms requires large labeled datasets, which are difficult to obtain. Synthetic images from deep generative models can mitigate data scarcity, but their effectiveness depends on fidelity to real images. Researchers often select models based on image quality metrics, prioritizing realism. However, the authors’ empirical analysis shows that high-fidelity, visually appealing synthetic images do not necessarily offer superior utility for downstream tasks; in one case, low-fidelity synthetic images outperformed high-fidelity ones. These findings emphasize the need for comprehensive analysis before integrating synthetic data into applications and highlight the potential value of low-fidelity synthetic images for training medical AI algorithms.