AI makes realistic rare gut lesion images from one photo to aid doctors and AI training

Study title: One-shot synthesis of rare gastrointestinal lesions improves diagnostic accuracy and clinical training

Researchers built EndoRare, a generative framework that creates varied examples of rare gastrointestinal lesions from a single reference image without retraining. It separates the features that define a lesion from incidental image attributes, keeping the first and varying the second.

Read the study abstract

Abstract Rare gastrointestinal lesions are infrequently encountered in routine endoscopy, limiting data for developing reliable artificial intelligence (AI) models and training novice clinicians. Here we present EndoRare, a one-shot, retraining-free generative framework that synthesizes diverse lesion exemplars from a single reference image. Using language-guided concept disentanglement, EndoRare separates pathognomonic lesion features from non-diagnostic attributes, encoding the former into a learnable prototyp…

By the Truvace deskOctober 27, 20261 min read1 sourceEvidence: Moderate · 68/100 — Controlled or validated comparison · 1 independent sourcePeer-reviewedTRV-2026-2284· cite & verify

One-shot synthesis of rare gastrointestinal lesions improves diagnostic accuracy and clinical training — "endoscopy" by Yuya Tamai is licensed under CC BY 2.0. To view a copy of this license, visit https://creativecommons.org/licenses/by/2.0/.

In brief

Rare lesions are seen so seldom in routine endoscopy that data for AI models and trainee exposure are scarce, so a data-efficient generator could help both. The reader study was exploratory and the authors call for prospective controlled validation before clinical use.

Main points

EndoRare generates diverse lesion images from one reference image and needs no retraining.
It uses language-guided concept disentanglement to separate diagnostic lesion features from non-diagnostic ones.
It was tested on calcifying fibrous tumor, juvenile polyposis syndrome, familial adenomatous polyposis and Peutz-Jeghers syndrome.
Novice endoscopists shown generated cases had improved image-level recognition in an exploratory reader study.

The gain

Synthetic images made from one reference photo were judged clinically plausible by experts, and adding them to training data raised AI detection rates for four rare gut conditions at low false-positive rates.

The rundown

The framework encodes the pathognomonic features of a lesion into a learnable prototype embedding, while varying other attributes so that the generated images are diverse rather than copies of the reference.

Validation covered four rare pathologies. Experts rated the synthetic images as clinically plausible, and using them to augment training data improved downstream classifiers, raising true positive rates at low false-positive rates.

In a reader study described as exploratory, novice endoscopists who were exposed to generated cases showed better image-level recognition. The authors present this as support for potential use in rare-lesion education.