Study: AI-generated prostate MRIs look real but risk misleading radiologists

A new paper from researchers Gabriel Paulo Maglalang Israel, Sol Gedde, and Alvaro Fernandez-Quilez — set to appear at SCAI 2026 — tackles a persistent bottleneck in prostate cancer care: MRI acquisition is costly, slow, and resource-intensive. The team built a 3D diffusion model that generates patient-personalized synthetic prostate MRIs conditioned only on routine clinical variables: age and PSA levels. Using a large public dataset, they first verified that these synthetic images were statistically comparable to real scanner-acquired MRIs using standard image similarity metrics. Then they ran a pilot, single-blinded, randomized study where a radiologist compared personalized synthetic MRIs against unconditioned synthetic versions, rating clinical plausibility, confidence, and the risk of misleading interpretation.

The results are a measured warning. The personalized synthetic MRIs were judged technically plausible, but expert opinion showed significant variability in confidence and flagged potential risks of misleading cues when age and PSA were embedded into image generation. In other words, making the MRI "personalized" introduced subtle patterns that could steer interpretation in the wrong direction. The authors note that while personalized synthesis may be technically feasible, such systems cannot be safely dropped into clinical workflows without deeper investigation into how generated features influence radiologist decisions. The study is a reminder that technical plausibility — as measured by similarity scores — does not equal clinical trustworthiness. It also highlights the gap between generating realistic images and generating images that support accurate medical judgment, especially as synthetic data tools move from research labs toward real-world trials.