EndoRare generates rare endoscopy examples from one reference image

Dates and assessment

Source published

2026-10-25

Bright published

2026-10-26

Substantive update

None recorded

Evidence state

Experimental

Independent verification

Not established by this source review

Last source review

2026-10-25

The claim in context

The human problem

Rare digestive-tract findings are difficult to learn when clinicians and AI developers have few examples.

The prior constraint

Relevant clinical examples can be scarce, limiting exposure and image-classifier training data.

AI’s actual role

EndoRare generates varied endoscopy examples from a single target reference image using a model already trained on routine medical images.

The documented result

One test asked whether synthetic examples helped an image classifier. Averaged equally across the four categories, its precision–recall area under the curve, or PR-AUC, was 0.708 with EndoRare, compared with 0.612 using Textual Inversion, a competing image-generation method. Supplementary Tables 9, 11, 13 and 15 PR-AUC summarizes the trade-off between finding positive cases and making correct positive predictions across different decision thresholds. The result does not mean the classifier was “70.8% accurate,” and it does not establish performance during a live examination. A separate exploratory study involved 10 novice endoscopists from four institutions, assessed on the same four conditions before and after exposure to generated examples. After pooling each reader’s decisions across those conditions, the authors reported an average F1 score rising from 0.581 to 0.901. F1 combines precision and recall. Accepted paper, Results and Table 1

Why it may matter

Generated examples could provide practice for uncommon findings, if benefits transfer to unfamiliar cases and persist.

Limitations

• For now, EndoRare offers a research direction for a familiar teaching problem: how to provide useful practice when examples are scarce. The evidence concerns static images, not fewer missed diagnoses or better patient outcomes.
• PR-AUC summarizes the trade-off between finding positive cases and making correct positive predictions across different decision thresholds. The result does not mean the classifier was “70.8% accurate,” and it does not establish performance during a live examination.
• That is an encouraging observation, but the readers took the same image test again after two weeks. They received no answers or performance feedback after the first test, and there was no comparison group. This design cannot separate the contribution of synthetic examples from familiarity with the test or practice. It leaves unanswered the question that matters most for education: does any improvement carry over to unfamiliar cases and last?
• Clinicians rated 700 generated outputs across seven methods. They did not know which method produced each image, but did know the images were synthetic.
• The classifier evaluation used five-fold cross-validation with the same splits across methods. Those folds were technical repeats within the evaluation, not five independent patient cohorts.
• The researchers have posted public research code, but its README states that no license has been declared and rights remain reserved. It should not be described as open source. The clinical images are not publicly available, and privacy protection has not been established. Data access and limitations

Original evidence

• EndoRare accepted paper · Nature Communications · paper
• EndoRare study · Nature Communications · paper
• Supplementary Methods 3.7 and Tables 9, 11, 13 and 15 · paper
• EndoRare public research code and licensing statement · repository

Attribution

Credit Bright AI Future and link the canonical Bright record.

• Link to the canonical Bright record.
• Keep material limitations with the claim they qualify.
• Link to the original evidence when repeating a substantive claim.
• Do not describe a source check or organization-reported result as independent verification.

Linked source material, quotations, trademarks and media remain subject to their owners’ terms. No reuse right is granted for third-party media.