Assessing realism of artificial intelligence-generated colorectal polyp images: International multicenter blinded reader study.

BACKGROUND AND STUDY AIMS: Prevention of colorectal cancer relies on detection and characterization of polyps during colonoscopy, yet access to large, shareable training datasets is limited. To address this issue, we developed a diffusion-based artificial intelligence (AI) model to generate synthetic polyp images in high-resolution and assessed their perceived realism in an international blinded reader study.

PATIENTS AND

METHODS

Fifty-three endoscopists from 46 centers across 14 countries evaluated 20 real and 20 AI-generated images in random order using our web platform Lutetia. Experts classified the images as either real or synthetic and rated their confidence. The primary endpoint was sensitivity of identifying synthetic images. Secondary endpoints were recognition of real images and overall accuracy. The trial was registered at clinicaltrials.gov with the identifier NCT07108569.

RESULTS

Sensitivity for detecting AI-generated images was 66% (95% confidence interval [CI] 65%-67%) with a specificity of 80% (95% CI 79%-81%); overall accuracy was 73% (95% CI 72%-73%). Low-confidence decisions were more frequent for AI-generated images and associated with longer annotation time ( P < 0.001).

CONCLUSIONS

Difficulties in differentiating synthetic from real images together with lower confidence and longer decision times shows that our diffusion model generates highly realistic polyp images.