Synthetic rare-lesion images helped five novice endoscopists—what remains unproven
A peer-reviewed study found better image-level recognition after novices trained on generated examples, but the same fixed test was used before and after and no patient outcomes were measured.
The short version
- What happened
- Five novice endoscopists improved on a fixed image-recognition test after seeing synthetic examples of four rare gastrointestinal lesions; the study did not test real procedures or patient outcomes.
- Who it affects
- Novice clinicians could gain more practice with conditions they rarely encounter, but realistic synthetic images can also create false confidence if training performance is mistaken for readiness to diagnose patients.
- What remains unproven
- The reader study included five novices, four conditions and the same fixed mixed image set before and after synthetic exposure, with no separate control group.
Keep in mind
The reader study included five novices, four conditions and the same fixed mixed image set before and after synthetic exposure, with no separate control group.
The classifier results came from curated image datasets and do not establish performance across hospitals, devices, prevalence levels or live procedures.
Two conditions showed ceiling effects, limiting the ability to detect improvement.
The approach reduces fine-tuning on patient pixels but has not undergone the formal privacy audits that the authors recommend.
Source analysis · 2 linked sources. Primary analysed source date: 25 October 2026. Earliest linked supporting source: 30 December 2025.
Read the full report · 4 min The evidence, methods, implications and limitations
Generated cases improved a narrow training exercise
Researchers led from Fudan University built EndoRare to generate varied endoscopy images from one real reference image for each of four uncommon conditions: calcifying fibrous tumour, juvenile polyposis syndrome, familial adenomatous polyposis and Peutz–Jeghers syndrome. Five novice endoscopists first studied one real anchor image per condition, took a fixed mixed-image test, then saw generated examples and repeated that same test. Overall recall rose by 0.400 and precision by 0.267. Gains were largest for the two less distinctive-looking conditions; performance on the other two was already near the ceiling.
That is useful evidence for an educational aid, not proof of safer diagnosis. A five-person before-and-after exercise can reflect familiarity with the repeated test, and image recognition is only one part of endoscopy. The study did not compare synthetic training with a separate control curriculum, follow retention over time or measure decisions during live procedures. Clinicians and training programmes should therefore treat the images as supplementary practice material, not substitute patients or validated competency assessment.[1][2]
The model results are promising, with important limits
In a blinded assessment of 700 randomly sampled images on a three-point scale, EndoRare received the highest overall rating, 2.32, compared with 2.09 for fine-tuned Stable Diffusion and 2.01 for DreamBooth; inter-rater agreement was moderate-to-substantial at Fleiss' kappa 0.59. A rare-lesion classifier trained with EndoRare augmentation reached macro precision–recall AUC 0.708, against 0.612 for Textual Inversion, and macro ROC-AUC 0.978. These are image-test metrics, not diagnostic accuracy in a hospital population.
The authors froze the diffusion backbone instead of fine-tuning it on the rare reference image, which may reduce direct copying. They explicitly say that this does not eliminate privacy risk and call for nearest-neighbour, membership-inference and attribute-leakage tests. Confidence would rise with a larger, multi-centre randomised reader study using unseen test cases, multiple devices and experience levels, retention follow-up, independent privacy testing and prospective evaluation during real clinical work.[1][2]
The distinction that matters
What the evidence actually follows
The paper tests three different steps; only the first two are model or image outcomes.
Generate
Create varied examples from one rare-lesion reference
Images rated for realism and class faithfulness
Train
Augment a classifier and expose five novices to synthetic cases
Higher image-test metrics in the reported experiments
Validate
Run controlled, prospective multi-centre studies
Still needed before claiming clinical benefit