Can we actually use chronic wound data to fake leprosy images?

Generating medical images for neglected diseases is usually a nightmare because there aren't enough photos to satisfy a hungry model. You can't just tell a diffusion model to "make a leprosy lesion" when the public datasets only have a few hundred images. This paper (arXiv:2609.13226v1) tries a workaround by using a massive amount of general chronic wound photography as a starting point to see if the "look" of a wound transfers over to leprosy.

The researchers built a three-stage pipeline to stop the model from just hallucinating random blobs.

How they handled the lack of masks

Since the wound datasets didn't come with masks, they used a DeepLabV3-ResNet50 segmentation network to generate them. This part actually held up reasonably well, hitting a validation Dice score of 0.876 and an IoU of 0.799. Without those masks, the next step would have been impossible.

Modifying Stable Diffusion 1.5

They didn't just plug images into a standard model. They built a mask-conditioned latent diffusion model using Stable Diffusion 1.5 components, but they had to rip open the UNet. They widened the input convolution from 4 channels to 11. This was necessary to shove in three mask feature maps and a blurred low-frequency context latent. They trained this beast on 3,280 region-of-interest wound crops before even touching the leprosy data.

Fine-tuning and the LPIPS reality check

Once the model understood what a generic wound looked like, they fine-tuned it on 708 leprosy image-mask pairs (sourced from 764 images across roughly 150 patients).

To see if the results were actually believable, they used LPIPS perceptual distance. They weren't just comparing generated images to real ones; they used a real-versus-real baseline on 242 anchor images to keep the math honest. Here is how the numbers shook out:

• Internal Perceptual Diversity (Generated): 0.662
• Internal Perceptual Diversity (Real Leprosy): 0.672 (95% CI [0.664, 0.680])
• LPIPS Distance from Real Distribution: 0.044

The diversity score shows there was no mode collapse—the model isn't just spitting out the same three "perfect" lesions. The generated images are technically measurably apart from the real ones, but they are within half of one standard deviation.

The funny part is that fine-tuning barely shifted the output distribution. The authors admit that the lesion geometry only reached the network through input concatenation, which is a fancy way of saying the model struggles with the specific shape of the lesion even if the texture looks right.

Basically, if you need synthetic leprosy images, starting with general chronic wounds works because the low-level "grossness" transfers well. The bottleneck isn't image quality anymore; it's getting the model to actually follow the semantic layout of the mask.