When Does Synthetic Realism Become Clinical Utility? Reflections on Generative Orthopantomogram Datasets

Dear Editor,

We read with great interest the article by Adnan et al.¹ on the development of synthetic orthopantomogram datasets using generative models. The study addresses a timely and important problem in dental artificial intelligence, namely the scarcity of shareable radiographic datasets and the need for privacy-preserving alternatives. The authors should be commended for comparing GAN- and diffusion-based approaches and, importantly, for evaluating generated images using both human observers and AI classifiers.

One aspect of the study that deserves further discussion is how 'realism' was operationalised. The human assessment was based on 40 images, of which only 20 were synthetic, and these were manually selected to include only high-quality outputs without visible distortions or blur. This is understandable in a pilot setting, but it also means that the reported inability of dentists to distinguish real from synthetic images reflects performance on a curated subset rather than on the full spectrum of generated images. In that sense, the findings may demonstrate perceptual plausibility under favourable sampling conditions more than robust realism across the broader synthetic distribution. Recent work in panoramic radiography likewise suggests that the value of synthetic image generation depends not only on visual fidelity, but also on how such images behave when integrated into downstream analytical tasks.²

A second point concerns the striking discrepancy between human and AI performance. While faculty and residents performed near chance, the AI models classified the same images with extremely high accuracy. The Grad-CAM analysis is especially revealing here, because the highlighted regions appear to be concentrated largely at the image periphery rather than over clearly relevant anatomical structures. This raises an important methodological question: where the models detecting subtle generator-specific signatures, pre-processing traces, or boundary artifacts rather than genuinely distinguishing anatomical authenticity? If so, the study may be showing that synthetic images are visually convincing to clinicians while still remaining computationally identifiable for reasons unrelated to clinical content. This distinction is important because recent evidence has shown that gradient-based saliency maps cannot always be assumed to provide trustworthy explanations of model reasoning in image-based diagnostic settings.³

A related issue is the paper’s forward-looking suggestion that diffusion-generated OPGs could be annotated and used to train diagnostic systems. This is an attractive proposition, but the present results do not yet establish clinical utility. Perceptual realism and a favourable FID score are not equivalent to preservation of diagnostically important features, class balance, or pathology-specific fidelity. Before synthetic OPGs can be considered reliable for downstream training, it would be valuable to know whether models trained with such data improve external validation performance, particularly for uncommon findings and heterogeneous acquisition settings.^2,4 More broadly, the challenge of generalizability remains one of the central barriers to clinical translation of AI models across institutions and imaging environments.⁴

We therefore believe this study makes a useful contribution not only by showing the promise of diffusion models in dental imaging, but also by highlighting the need to distinguish visual realism, algorithmic detectability, and task-level usefulness. That distinction may become increasingly important as synthetic data moves from technical feasibility to clinical research and, eventually, to implementation in oral healthcare.