Gastroenterologists Struggle to Detect AI-Generated Colonoscopy Images
In a study evaluating the detectability of artificial intelligence (AI)–generated content, practicing gastroenterologists were only slightly better than chance at distinguishing real from synthetic colonoscopy images. The findings were published in Biomedicines
“Synthetic and pseudosynthetic images can be used to extend colonoscopy datasets, which, in turn, are used to train AI-detection models, yet their clinical acceptability depends on whether medical professionals can still recognize non-real content,” explained the study authors. “If synthetic or pseudosynthetic images are perceptually indistinguishable from real ones, they may serve not only as valid training inputs for AI but also as educational tools and validation resources in medical practice.
The investigators evaluated the ability of practicing gastroenterologists to discriminate real, pseudosynthetic, and synthetic polyp images and to determine how training level and synthesis method affect detection.
A total of 32 gastroenterologists, 18 residents and 14 seniors, reviewed and classified 24 polyp images, including 8 real, 8 augmented, and 8 generated with two different types of AI algorithms (CycleGAN, n=4; diffusion, n=4), as “real” or “non-real”.
Overall, the gastroenterologists correctly classified the images 61.2% of the time (95% CI, 57.7%-64.6%), with similar performance between resident and senior gastroenterologists (62.3% vs 59.8%; P=0.54). Agreement between gastroenterologists was fair (κ=0.30; 95% CI, 0.15-0.43), reflecting widespread uncertainty among the reviewers.
When participants distinguished real from non-real images, sensitivity and specificity were 70.7% and 78.5%, respectively. Accuracy varied by image type: 70.7% sensitivity for real images, 51.6% for augmented, and 61.3% for synthetic. Notably, all CycleGAN-generated images were identified as synthetic (128/128), but diffusion-generated images were correctly identified only 22.7% of the time (P<0.001).
“Clinicians detect non-real colonoscopy images only slightly above chance, irrespective of experience,” concluded the authors. “The diffusion synthesis method creates images that escape human scrutiny, suggesting the need for automated authenticity safeguards before synthetic datasets are applied in clinical or AI-validation contexts.”
The investigators believe that highly realistic synthetic and pseudosynthetic polyp images can be used to enhance AI training datasets for colonoscopy. Because these images are often indistinguishable from real ones, they could help AI systems better detect subtle or atypical lesions. This strategy may reduce adenoma miss rates and improve colorectal cancer screening outcomes by strengthening the accuracy and generalizability of AI-assisted detection.
Ioanovici AC, et al. Biomedicines. 2025;13(7):1561. doi:10.3390/biomedicines13071561
SLEEP 2026 Recap: The New Era of Integrated Sleep Medicine
Q&A: Does Migraine Increase Risk for Cervical Artery Dissection?
Protocol for Outpatient Management of Transplant Recipients With Delayed Graft Function
180 Mount Airy Rd., Suite 205
Basking Ridge, NJ 07920
About Us
© 2026 Formedics, LLC. All Rights Reserved.
Leave a Reply