THIS CONTENT IS BROUGHT TO YOU BY NTNU Norwegian University of Science and Technology - read more
AI helps doctors detect intestinal disease
Can you tell which of the images below is real? Polyps in the intestine are common, but they are often removed to prevent them from developing into bowel cancer. AI can identify the polyps that may be harmful.
Published
Polyps in the intestine are usually benign growths on the inside of the colon or rectum. Many people have them. Nevertheless, it is commom to remove them because some can develop into bowel cancer over time.
If diseases of the stomach and intestine are suspected, patients are typically examined using a long thin tube and a camera. The equipment is inserted into the throat (gastroscopy) or the rectum (colonoscopy).
Artificial intelligence is now helping doctors prioritise which patients require a gastro- or colonoscopy.
Capsule endoscopy
An alternative option is to use a capsule with a built-in camera. The patient swallows the capsule. Seven hours later, the camera has taken tens of thousands of images from the inside of the stomach and intestine.
“Capsule endoscopy uses a small camera that passes through the gastrointestinal system. The camera is inside a large pill that the patient swallows, and which later comes out in the toilet bowl,” explains Øistein Hovde.
He is a senior consultant at the Gastroenterology Laboratory at Gjøvik Hospital.
The laboratory receives patients who need to be examined and treated for gastrointestinal diseases.
Up to 50,000 images from the intestine
The problem? The amount of data. A single capsule can take up to 50,000 images during one examination. The journey takes six to seven hours. The camera continuously sends images to a computer the patient carries with them.
“This is simply a lot of data for a doctor to go through manually, even though the images can be run at high speed,” says Marius Pedersen.
He is a professor at Colourlab, part of NTNU's Department of Computer Science.
Researchers at this laboratory and doctors and researchers at the hospital in Gjøvik have now collaborated to solve precisely this problem: finding a faster and reliable way to identify the images doctors need in order to detect disease.
More images are needed
The starting point was 1,350 images from 10 patients with the inflammatory bowel disease ulcerative colitis. These images were captured using a colonoscopy.
The same patients also underwent a capsule endoscopy.
However, the AI needed to be trained before it could identify abnormal polyps captured by the capsule camera.
But for AI to be able to identify images of diseased polyps from the images taken by the capsule cameras, it needed training. And that required even more images than the 1,350 they already had.
“Artificial intelligence needs to have large amounts of data to be effective. At the same time, it was difficult to collect enough traditional images using manual tubes and cameras, due to time comstraints and privacy concerns,” says Pedersen.
The solution was to create more images artificially.
‘Doctored’ images
The research team used diffusion models and the machine-learning method CLIP to generate new, realistic images of polyps.
The content of the images had to be varied. For example:
Where in the intestine the image appeared to be taken
The number of polyps
The location of the polyps
Visibility conditions that could be affected by debris, blood, or bubbles
Polyps at different stages of development
Can doctors tell the difference?
In addition to first obtaining images manually from the patients using colonoscopy, the researchers also used the pill cameras on the same patients. These were used in the test.
Would gastroenterologists be able to tell the difference between the real and artificially generated images?
“The doctors answered correctly in 54 per cent of cases. The artificial images were difficult to distinguish from the real ones,” says Pedersen.
Experienced and less experienced doctors performed about equally well.
“This tells us that the synthetic images in this context are of comparable quality to conventional images,” says Pedersen.
Less discomfort for the patient
The goal is not to replace current examinations, but to make the process easier and more efficient.
“Colonoscopy remains the gold standard. It allows us to both examine the inside of the intestine and stomach, and to take samples,” says Hovde.
One goal of the new AI-created images is for assessments to be more consistent from doctor to doctor. Today, two specialists may interpret the same image differently.
“With more and better images, it can be easier to train doctors and reach a common agreement on what we are actually seeing,” says Hovde.
Easier to prioritise
Hovde also believes AI can help doctors prioritise which images to examine.
Instead of browsing through tens of thousands of images, a doctor could, for example, be presented with the hundred most relevant pictures. They would also receive precise information about where in the intestine the images were taken.
“AI would reduce the number of examinations and require less analysis work, which often takes three to four days per patient. With this approach, it could be reduced to one day. The pill camera also reduces patient discomfort,” says Hovde.
Useful for screening
The researchers believe the pill camera could also be useful in detecting bowel cancer.
Norway has one of the highest rates of the disease in the world, with around 5,000 people diagnosed each year. Excluding gender-specific types of cancer, such as ovarian and prostate cancer, it is the most common form of cancer.
Norway has a national screening programme for bowel cancer. All 55-year-olds are invited to submit a stool sample, which is tested for traces of blood.
Many bowel cancer cases start as small polyps in the intestine. At least nine out of ten polyps will never develop into cancer. But size and appearance can provide an indication of cancer risk.
“Pill cameras could be used for patients whose stool samples indicate that further investigation is needed,” says Hovde.
He explains that if the images show polyps whose size and appearance suggest a risk of cancer, the patient should undergo a colonoscopy.
But if traces of blood are found in the stool and the pill camera reveals only harmless polyps, the patient could avoid a colonoscopy.
This will spare pateints discomfort and reduce costs for the healthcare system.
Better battery and optics
Hovde and Pedersen highlight the importance of the collaboration between the hospital and NTNU.
“The results are a good example of how clinical experience and technological research can complement one other,” says Pedersen.
In the collage at the top of the article, the artificial image is on the left and the real one is on the right. Were you able to tell the difference?
Reference:
Xu et al. Controllable Image Synthesis for Endoscopy: Leveraging Text and Spatial Guidance in Diffusion Models (PDF), Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) Workshops, 2026.
———
Read the Norwegian version of this article on forskning.no
More content from NTNU:
-
Why making lighter and better batteries takes time
-
Professor: Why we should expect cyberattacks during major events
-
Migraines leave traces throughout the body that researchers can now detect
-
Insect expert hunts for unknown life in New York
-
No, your family relationships don’t get worse even if you spend a lot of time on social media
-
Can artificial intelligence help save nature?
