Doctors can now test AI diagnosis tools on rare genetic diseases — using real patient photos plus AI-generated ones

Researchers created the first benchmark dataset of pediatric facial images for rare genetic diseases: 456 real photos across 103 conditions, paired with AI-generated synthetic images filtered to match real phenotypes. This lets AI researchers build and test diagnostic tools in the extreme low-data environment where rare diseases actually live — instead of pretending they have thousands of examples.

Until now, AI diagnosis tools for rare genetic diseases couldn't be properly tested because no curated dataset existed. Researchers either built private datasets (not comparable across studies) or trained on synthetic data that didn't match real patient variation. This dataset breaks that logjam. It means the next generation of rare disease screening tools can be evaluated on real constraints — 4 to 5 examples per condition — instead of being benchmarked on fantasy datasets with thousands of images. The synthetic augmentation framework they tested (filtering generated images by facial landmark similarity) also shows that you can stretch thin datasets without hallucinating phenotypes away. That matters because it's the first time anyone has measured whether synthetic medical imagery actually preserves diagnostic validity, not just visual plausibility.