Artificial intelligence is transforming medical imaging, but there's a catch: AI systems need thousands of labeled images to learn effectively, and rare diseases don't generate enough data. Now, researchers at the Norwegian University of Science and Technology (NTNU) in Gjøvik have demonstrated a potential workaround—creating synthetic endoscopy images that are so realistic, gastroenterologists can barely distinguish them from the real thing.

In a validation test, specialist physicians correctly identified real versus AI-generated endoscopy images only 54% of the time—essentially chance-level performance. This suggests the synthetic images have crossed a critical threshold of visual authenticity, though the researchers emphasize that larger clinical validation studies are needed before these training methods can be deployed in actual patient care.

The Data Scarcity Challenge in Gastroenterology AI

The problem these researchers are tackling is substantial. A single capsule endoscopy examination—where patients swallow a pill-sized camera that travels through the digestive system—can generate up to 50,000 images over 6-7 hours. Reviewing this massive volume of images is time-consuming for physicians, making it an ideal candidate for AI assistance. However, training AI systems to recognize abnormalities requires large datasets of labeled images, particularly for conditions like ulcerative colitis.

The NTNU team started with only 1,350 colonoscopy images from 10 patients with ulcerative colitis—far too few for robust AI training. Using generative AI, they created synthetic images that could expand this limited dataset. Importantly, their system allows researchers to control specific variables including polyp location, disease stage, and viewing conditions such as the presence of blood, bubbles, or bowel contents.

From Image Generation to Clinical Application

The research collaboration between NTNU Gjøvik and Gjøvik Hospital, involving researchers Lu Xu, Anuja Vats, Marius Pedersen, and Kiran Raja, has produced work published in the Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision Workshops, 2026.

The AI system is designed to pre-sort images and flag potentially abnormal findings for physician review, not to replace clinical judgment. This approach could significantly reduce the time physicians spend reviewing capsule endoscopy images by pre-filtering the most clinically relevant frames.

Capsule endoscopy offers a non-invasive alternative to traditional colonoscopy, though it cannot perform biopsies or remove polyps. If AI-assisted analysis proves clinically valid, it could make capsule endoscopy more practical for screening and surveillance, potentially offering patients a more comfortable initial screening option. However, high-risk findings would still require traditional colonoscopy for tissue sampling and therapeutic intervention.

The Path Forward

While the inability of gastroenterologists to reliably distinguish synthetic from real images is encouraging for training purposes, the researchers acknowledge that clinical validation remains the critical next step. The technology remains investigational, and regulatory pathways for using synthetic data in medical AI development are still unclear.

For medical AI companies developing computer vision tools for gastroenterology, this approach could address a fundamental bottleneck—the scarcity of labeled training data, particularly for rare conditions. The ability to generate controlled, diverse synthetic images may accelerate development cycles and improve model performance across varied patient populations, though the path from research demonstration to clinical deployment requires rigorous validation.

Sources

  1. AI Images Could Help Doctors Find Bowel Disease Faster — Knowridge Science Report