Fudan-Led AI Builds Rare Gut-Lesion Training Images From One Photo

A team led by Fudan University in Shanghai has built an artificial intelligence model that generates varied images of rare gastrointestinal lesions from a single example photo, to help train both diagnostic software and junior doctors. The paper, published in Nature Communications on 25 October, describes EndoRare as “a one-shot, retraining-free generative framework that synthesizes diverse lesion exemplars from a single reference image”.

The authors, from Fudan, Zhongshan Hospital, Zhejiang University, Imperial College London and the University of Manchester, tested it on four rare conditions. They report that synthetic images improved AI classifiers and that, in a small exploratory study, novice endoscopists who studied the generated cases recognised rare lesions better afterwards. The authors stop short of claiming it improves real-world diagnosis, and it is not in clinical use.

Rare lesions leave doctors and software short of examples

The problem the paper sets out is one of numbers. “Rare gastrointestinal lesions are infrequently encountered in routine endoscopy, limiting data for developing reliable artificial intelligence (AI) models and training novice clinicians,” the abstract says. A trainee may see few or no examples of some conditions during training, and a classifier cannot learn well from a handful of images.

The four conditions tested were calcifying fibrous tumour, juvenile polyposis syndrome, familial adenomatous polyposis and Peutz-Jeghers syndrome. The image data came from routinely collected records, and the study was approved by the ethics committee of Zhongshan Hospital, Fudan University, according to the manuscript.

How EndoRare separates a lesion from its background

According to the abstract, EndoRare uses “language-guided concept disentanglement” to separate “pathognomonic lesion features from non-diagnostic attributes”. The features that define the lesion are encoded “into a learnable prototype embedding”, while the other attributes are varied “to ensure diversity”. In plain terms, the model tries to keep what makes a lesion recognisable and change everything else, such as colour, lighting and surrounding tissue, so one photo yields many different but faithful examples.

The paper reports two results from this. “Experts judge the synthetic images clinically plausible, and data augmentation significantly improves downstream AI classifiers, increasing true positive rates at low false-positive rates,” the abstract says. The authors add, in the manuscript, that these classifier scores “should not be interpreted as sufficient evidence for clinical deployment”, because rare conditions mean even a small false-alarm rate can produce many unnecessary alerts. They suggest such classifiers may be more plausibly explored as “second-reader, triage, or case-prioritization aids” than as “autonomous first readers”.

The reader study was small and had no control group

The human test involved ten novice endoscopists from four hospitals, according to the manuscript. Each took an image-recognition test, then studied EndoRare-generated cases, then took the same test again two weeks later with no feedback in between. Pooled across the four conditions, their F1 score, a combined measure of accuracy, rose from 0.581 to 0.901, the paper reports.

The abstract words this carefully: “In an exploratory reader study, exposure to EndoRare-generated cases is associated with improved image-level recognition by novice endoscopists.” The authors write that because the same images were reused, “residual familiarity with repeated images and general learning effects cannot be fully excluded”. All tests used still images rather than live procedures.

The peer review file shows how the study changed. One reviewer described the first version’s reader study as “small (n=5), single-centre” and asked for the gains to be framed “as preliminary and hypothesis-generating”. In their reply the authors said they had expanded it to ten readers at several sites and now present the results “as preliminary image-level evidence rather than definitive evidence of educational efficacy”.

The code is posted on GitHub. The patient images are available only on request for non-commercial research, the manuscript says.

TSN has covered other recent Nature Communications health AI papers, including King’s College London’s HRDPath model for ovarian cancer slides and a Johns Hopkins study of AI eye exams for young people with diabetes.

Synthetic cases help most where real ones are scarcest

The appeal of this approach is that it targets the place where both trainees and algorithms are weakest. Common conditions already generate plenty of training material; it is the rare ones, seen a few times in a career, that are easiest to miss. A tool that can turn one archived image into a teaching set could, in principle, give a junior doctor exposure that would otherwise take years, and give a classifier more than a handful of examples to learn from.

That is also why the classifier result matters more than it might seem. Gains at low false-positive rates are the relevant ones for rare disease, because a screening aid that cries wolf on every polyp would be switched off. The authors themselves point to second-reader or triage uses rather than replacing the endoscopist, which is a more modest and more credible role than “AI diagnoses rare disease”. Compared with tools already in the field, such as the TB X-ray screening Google described this week, EndoRare is at a much earlier stage.

The training claim rests on one before-and-after test

The education result is the headline-friendly part, and the weakest. Ten readers, the same test taken twice, and no group that studied real images instead: that design cannot separate the effect of EndoRare from simply having seen the test before or having had two more weeks of practice. The authors say so, and the reviewers pushed them to say so more clearly.

The useful next step is the one the paper names. A controlled trial, with separate test sets and a comparison group trained on conventional material, would show whether synthetic cases teach anything beyond what real cases already provide. Until then, EndoRare is best read as a promising data tool with an encouraging but unproven teaching signal. The test that matters is whether a trainee who studied its images spots a real Peutz-Jeghers polyp during a live procedure.

Sources

• Nature Communications: One-shot synthesis of rare gastrointestinal lesions improves diagnostic accuracy and clinical training (Yu, Zhu, Fu et al., published 25 October 2026), including the Transparent Peer Review file and manuscript source supplied as supplementary information
• GitHub: EndoRare code repository

Related stories

• KCL AI Spots DNA-Repair Defect in Ovarian Cancer From Routine Slides
• AI Eye Exams Closed a Youth Diabetes Screening Gap, Study Says
• Google Says Partner’s AI Has Read 25,000 X-Rays in TB Screening