AI-Driven Imaging Transforms IBD Patient Outcomes

Researchers from the Schubert-Martin Inflammatory Bowel Disease (IBD) Center at Cincinnati Children’s are developing predictive algorithms to help improve care for patients with Crohn’s disease, ulcerative colitis and very-early-onset IBD. Jazz Dhaliwal, MBBS, MRCPCH, MSc, a physician scientist trained in IBD, leads a team working on advanced computational research using imaging data.

“IBD lends itself to the imaging domain because we undertake endoscopy to evaluate the colonic mucosa, taking biopsies which are then subsequently reviewed by our pathologists,” she explains. “We also undertake MRI, specifically MRE, to evaluate the small bowel as standard of care. Thus, there is an opportunity to apply advanced computational approaches to evaluate clinical imaging data to derive and answer pertinent questions that are important for our patient cohort.”

The Algorithm-Based Research

Dhaliwal’s team has completed several studies, with more in progress. For their research, they use pediatric cohort data from large, rare disease research studies—a requirement for any kind of artificial intelligence (AI) work or machine learning.

“What we’ve done, with our pathology work for example, is to develop a predictive algorithm from baseline mucosal tissue samples using an agnostic approach,” Dhaliwal says. “It’s not driven by a pathologist’s assessment whereby variability is not uncommon.”

As part of the PROTECT study, the team was able to evaluate digitized mucosal specimens at diagnosis, prior to therapy. They applied machine learning algorithms to extract handcrafted or histomic features. These features look at shape, color, spatial orientation and morphology. Viewing the image and underlying variation, they then predicted with close to 0.9 accuracy if and how pediatric ulcerative colitis would progress.

“We derived a natural language processing algorithm to predict from endoscopy reports disease features and disease location,” she says. “We have applied our algorithm to automate endoscopic disease severity quantification, and we’re just now writing this up for the public. Endoscopy images and deep phenotyping are in their infancy, but there’s more to come.”

Predicting Outcomes to Ease Burden

The predictive algorithms have many benefits for the field of IBD:

• Eases the burden of pathologists in a clinical setting: Machine and deep-learning approaches can identify standard histological features to help pathologists when extracting disease features from pathology and endoscopy reports. While algorithms don’t replace the pathology, they can make it faster and easier on the pathologist. Dhaliwal refers to it as “calling a friend.” For example, with radiology, a prompt appears that is the algorithm for what a provider is seeing. It’s a friend offering support, except the friend is within the developed system.
  
  
• Helps with therapy stratification: Upfront, providers may have four or five different IBD therapy options, plus more coming to market. With data support, there’s more of a starting point for which drug will ultimately benefit a patient the most.
  
  
• Reassures patients: After diagnosis, providers can have conversations with patients that may be more reassuring. Algorithms back up predictions that “you’ll probably do well on this medication.”

Large Datasets Propel the Research

The IBD Center at Cincinnati Children’s is unique because it has access to large datasets, as well as institutional support for innovative programs, such as looking at AI to improve IBD treatment.

“This is a very exciting field, but at present, there are not a lot of physician scientists focusing on pediatric IBD,” Diwali says. “At Cincinnati Children’s, there’s a philosophy of collaborative research and working together. We have developed a collaborative team with expertise in data management, computer science, biostatistics and IBD pathology. Over the last few years, we have been able to build upon this strong infrastructure.”

As the work continues, the team also has an interest in social determinants of health and is aware of ethnoracial variations in existing datasets. They’ve stressed the importance of considering potential biases when developing algorithms. And that medical practice changes over time.

“It’s important because you can develop these innovative approaches, but we are very conscious of not further exacerbating the bias,” Dhaliwal emphasizes. “There will be variation over time, largely because practices evolve, and populations may change—these potential factors need to be taken into account if algorithms are deployed in clinical practice. Updating the algorithm to adjust for change would be imperative.”

(Published October 2024)

Read More from Gastroenterology, Hepatology and Nutrition