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

Area of Science:

• Medical imaging
• Artificial intelligence in healthcare
• Gastroenterology

Background:

• Colorectal cancer (CRC) prevention depends on accurate polyp detection during colonoscopy.
• Limited availability of large, shareable training datasets hinders AI development for polyp analysis.
• High-resolution synthetic polyp image generation is crucial for advancing AI in CRC screening.

Purpose of the Study:

• To develop a diffusion-based AI model for generating high-resolution synthetic colon polyp images.
• To assess the perceived realism of AI-generated synthetic polyp images through an international blinded reader study.

Main Methods:

• Fifty-three endoscopists from 46 international centers evaluated 40 images (20 real, 20 synthetic) on the Lutetia web platform.
• Experts classified images as real or synthetic and reported confidence levels.
• Primary endpoint: sensitivity in identifying synthetic images; Secondary endpoints: recognition of real images and overall accuracy.

Main Results:

• The AI model achieved 66% sensitivity and 80% specificity in identifying synthetic polyp images.
• Overall accuracy in distinguishing real from synthetic images was 73%.
• Lower confidence and longer annotation times were associated with AI-generated images, indicating their high realism.

Conclusions:

• The diffusion-based AI model successfully generates highly realistic synthetic colon polyp images.
• The study highlights the potential of AI-generated images to overcome data limitations in CRC research.
• Expert difficulty in differentiating synthetic from real images validates the model's effectiveness.