Amii Fellow and Canada CIFAR AI Chair Ross Mitchell and a team at the University of Alberta have secured a new $1-million grant from Canadian Institutes of Health Research (CIHR) to develop and test new AI tools that could help physicians diagnose and monitor inflammatory bowel diseases (IBD), like Crohn’s disease.

It’s a project that Mitchell says will advance both machine learning in healthcare as well as our understanding of IBD.

“This medical need prompted a computer science advance, which in turn fed back and will create a medical advance,” he says.

Approximately 322,600 Canadians are living with some form of IBD, according to a  2023 report by Crohn’s and Colitis Canada. And that number is growing rapidly — IBD is estimated to affect around 1.1% of the country (470,000 people) by 2035.

Inflammatory bowel disease has a significant impact on the lives of those suffering from it. It is also extremely difficult to diagnose and monitor accurately. Even severe symptoms can be vague and difficult to articulate to a doctor, with a lot of overlap with other intestinal conditions.  

The progression of IBD can be monitored through a CT scan, but it is a slow and imperfect process. Inflammation can often be missed by scans due to obstruction and the length of the bowel which ranges from 4.5 to 6 meters in length in an adult. Even when the inflammation is picked up in a CT, it can be difficult for a trained radiologist to spot by eye alone.

The team at the University of Alberta secured the grant to improve this process. Made up of experts in computer science, radiology, epidemiology and gastroenterology, the team is using thousands of patient scans to build and validate an AI model that can automatically find the boundaries of the intestinal tract  in CT scans of IBD sufferers. This process, called “segmentation”, is the first step to Mitchell’s ultimate goal: identifying specific medical indicators, called biomarkers, that could help physicians track the progression of a person’s disease, leading to better treatment.

“Right now we don't have a quantitative biomarker. We can't say, ‘Oh, last year at this time, 20% of your bowel was involved, and this year it's 25%.’ Therefore, your disease has gotten worse, and we're going to switch your treatment to Drug B from Drug A,” he says.

To build this tool, Mitchell says they have access to 11,208 specialized abdominal CT scans from 105 Alberta hospitals, pulled from a large dataset of scans taken from more than 70,000 Albertan IBD patients. (Mitchell used the same dataset for a project last year, finding evidence of a link between IBD diagnoses and anxiety/depression.)

Combined with other scans from public databases, the researchers are working with more than 20,000 images — the largest dataset ever assembled for AI  intestinal segmentation.

But Mitchell is quick to point out that it isn’t just the size of the dataset that matters, but also the quality.

Or rather, the lack of quality.

Mitchell says that most publicly available datasets for training models on intestinal images come from clinical trials. Those are usually very high quality, meticulously positioned scans of healthy bowels.  But inflamed intestines look different than healthy ones. And the extremely high-quality scans done during clinical trials aren’t the same as the faster, often less clear images a patient would get taken in an imaging lab or hospital. Training a model on the types of images that are actually seen by a radiologist with an IBD patient means it is more likely to be able to do that segmentation properly.

The CIHR grant, “Interdisciplinary AI development for clinical translation of intestinal tract segmentation in inflammatory bowel disease,” provides $1,040,400 in funding over five years for the team to build the tool and validate its efficiency. Mitchell says the grant was accepted on its first application, a rarity for CIHR grants, which often take several rounds of application over 2-3 years.  It was one of 39 Alberta grants awarded by the CIHR, out of a pool of more than 315. He credits its quick acceptance both to the impact that better IBD monitoring could have for hundreds of thousands of Canadians, as well as the strong potential that AI has for making medical imaging more efficient and providing more access to underserved populations.