AI Generates Rare Gut Lesion Images From a Single Photo to Sharpen Diagnostics

Rare gastrointestinal lesions are so infrequently encountered in routine endoscopy that clinicians in training may go years without seeing one, and the scarcity of real images has likewise limited the data available for developing reliable artificial intelligence models. Reporting in Nature Communications, a research team has now presented EndoRare, a one-shot, retraining-free generative framework that synthesizes diverse lesion exemplars from a single reference image. The technical core of the approach is language-guided concept disentanglement.

EndoRare separates the pathognomonic features that define a lesion from non-diagnostic attributes, encoding the defining features into a learnable prototype embedding. By varying only the non-diagnostic attributes across generated images, the system produces visually diverse exemplars while preserving the diagnostic content that matters for recognition, all without retraining the underlying model. The researchers validated the framework across four rare pathologies: calcifying fibrous tumor, juvenile polyposis syndrome, familial adenomatous polyposis, and Peutz-Jeghers syndrome.

An important checkpoint for any generative tool intended for medical use is whether its outputs look credible to specialists. Expert reviewers judged the EndoRare-generated images to be clinically plausible, indicating that the synthetic exemplars retained the features that define each pathology rather than producing superficially convincing but diagnostically empty images. The synthetic data also delivered measurable gains for machine learning.

When used to augment training sets, EndoRare-generated images significantly improved downstream AI classifiers, increasing true positive rates at low false-positive rates. Beyond algorithms, the team ran an exploratory reader study with novice endoscopists. Exposure to EndoRare-generated cases was associated with improved image-level recognition of the rare lesions, pointing to a role for synthetic exemplars in clinical education where real cases are too scarce to build a training curriculum around.

The authors are careful about scope. The reader study was exploratory, and the association between exposure to synthetic cases and improved recognition does not by itself establish a causal training benefit. The findings support EndoRare's potential as a data-efficient tool for rare-lesion education while motivating prospective controlled validation before any conclusions about deployment are drawn.

Even with those limitations, the work demonstrates that a single reference image can be stretched into a diverse, expert-endorsed set of training exemplars for pathologies that are rarely photographed. For rare gastrointestinal lesions, where both AI development and novice clinician training have been constrained by the same shortage of real cases, EndoRare offers a data-efficient way to address both needs from the same synthesized resource, pending the controlled studies that will determine how well the approach performs in practice.

Subject of Research: One-shot synthesis of rare gastrointestinal lesions improves diagnostic accuracy and clinical training

Article Title: One-shot synthesis of rare gastrointestinal lesions improves diagnostic accuracy and clinical training

Article References: Yu, J., Zhu, Y., Fu, P., Chen, T., Wang, Z., Wu, F., Li, Q., Zhou, P., Wang, S., & Yang, X. (2026). One-shot synthesis of rare gastrointestinal lesions improves diagnostic accuracy and clinical training. Nature Communications. https://doi.org/10.1038/s41467-026-78492-y

Image Credits: AI Generated

DOI: 10.1038/s41467-026-78492-y

Keywords: EndoRare, gastrointestinal endoscopy, rare lesions, generative AI, image synthesis, data augmentation, clinical training, Peutz-Jeghers syndrome, familial adenomatous polyposis, One-shot, synthesis, rare