AI Can Now Fake Leprosy Photos to Help Doctors Spot It Faster
Rare diseases lack training photos. This AI makes convincing stand-ins.
A researcher built an AI image generator from Stable Diffusion 1.5 components to synthesize leprosy lesion images, aiming at a data problem: public annotated leprosy image sets number in the hundreds, far below what generative models require. The model was first trained on 3,280 region-of-interest crops from chronic wound photography, then fine-tuned on 708 leprosy image-mask pairs drawn from 764 images of approximately 150 patients. The generated images showed no mode collapse — their internal perceptual diversity was statistically indistinguishable from the real leprosy set — and sat 0.044 LPIPS outside the real distribution, measurably apart but under half a standard deviation. The paper concludes chronic wound photography is a viable donor domain, with the remaining barrier being semantic control rather than image quality.
- Only about 764 real leprosy photos exist publicly — too few to train diagnostic AI, so this researcher generated thousands of fakes instead
- The AI first learned from 3,280 ordinary wound photos, then adapted to leprosy using just 708 labeled examples
- The synthetic images are visually close to real ones and varied enough to be useful, but the AI copies outlines rather than truly understanding the disease
Why It Matters
Rare diseases could get diagnostic AI years sooner by borrowing data from common conditions instead of waiting for scarce photos.