A NOVEL AI approach called EndoRare improved the detection of rare gastrointestinal lesions by AI models and helped novice endoscopists recognise these conditions more accurately in a study. The approach creates synthetic endoscopic images from a single reference image, potentially addressing a key challenge in developing diagnostic tools and training clinicians: the limited availability of images of rare conditions.
Rare gastrointestinal lesions can be difficult to study because relatively few examples are available for training AI systems or educating clinicians. EndoRare was designed to generate additional images while preserving the features that help distinguish one condition from another, without requiring the AI model to be retrained.
The researchers evaluated images representing four conditions: calcifying fibrous tumour, juvenile polyposis syndrome, familial adenomatous polyposis, and Peutz-Jeghers syndrome. Experts assessed 700 generated images for realism, accuracy in representing the intended condition, and whether they were meaningfully different from the original images. EndoRare received a mean overall score of 2.32 out of 3.
When synthetic images were used to support AI classification, EndoRare outperformed the strongest comparison method on several measures. Its area under the precision-recall curve was 0.708, compared with 0.612 for the strongest alternative. It also achieved a higher area under the receiver operating characteristic curve, at 0.978 versus 0.961.
At a false positive rate of 20%, EndoRare achieved a true positive rate of 98.8%, compared with 90.0% using conventional image augmentation. These findings suggest that synthetic images could help AI systems identify rare lesions while reducing the risk of missed cases under the tested conditions.
The researchers also assessed 10 novice endoscopists before and after training with synthetic images. Their ability to identify rare lesions improved: mean recall increased from 55.0% to 93.8%, while precision rose from 62.3% to 87.0%. Improvements were particularly notable for juvenile polyposis syndrome and calcifying fibrous tumour.
However, the study used static images rather than live endoscopy, and participants saw the same image set before and after training. Familiarity with the images may therefore have contributed to the improvement.
The authors emphasised the need for larger studies across multiple centres, including assessments using endoscopy videos and real-world clinical settings, before the approach’s benefits can be established in practice.
Reference
Yu J et al. One-shot synthesis of rare gastrointestinal lesions improves diagnostic accuracy and clinical training. Nat Commun. 2026;DOI: https://doi.org/10.1038/s41467-026-78492-y.
Featured image: Images for you on Adobe Stock.
Adobe Stock
Each article is made available under the terms of the Creative Commons Attribution-Non Commercial 4.0 License.
