Using Computer Vision to Improve the Evaluation of Dysplasia in Inflammatory Bowel Disease

About This Grant

This study aims to develop new methods for detecting pre-cancerous dysplasia on colonoscopy and histology in patients with inflammatory bowel disease (IBD). IBD is associated with a higher incidence of colorectal cancer compared to the general population. However, IBD dysplasia is more difficult to detect on colonoscopy because lesions are flat, irregular in shape, and coincide with inflammation. In efforts to combat visualization problems, most gastroenterologists continue to perform random mucosal biopsy for increased sensitivity of dysplasia detection on colonoscopy. Accessory measures to help enhance dysplasia detection including high- definition endoscopy, dye chromoendoscopy, and narrow band imaging require extensive expertise, increase procedure duration, and have not been definitively shown to improve dysplasia detection rates. In addition to difficulty detecting dysplasia on colonoscopy, pathologists face similar ambiguity when evaluating dozens of biopsies provided from every colonoscopy. Beyond reviewer fatigue, pathologists are challenged to separate inflammation from dysplasia and the grade of severity, typically requiring referral to experts at high volume centers for second opinion review. Machine learning and computer vision methods are well suited to address clinician limitations in detecting visual features of IBD-related colonic dysplasia. Our multi-disciplinary team’s prior work developing methods to improve endoscopic disease activity assessments and quantify histologic imaging using machine learning will be adapted and applied to dysplasia detection in this proposed project. We will pursue three aims to achieve our goal of determining whether computer vision models can match or exceed the diagnostic ability of experts for detecting dysplasia on colonoscopy and histology. Aim 1 will build computer vision models trained to infer histologic ground truth using endoscopic imaging for detecting the presence of dysplasia on standard colonoscopy video from multiple centers. Methods will incorporate both still image classifier pipelines and new generative diffusion-based model architectures for full video analysis. Aim 2 will evaluate the performance of both experts and new FDA-approved AI assistant tools in colonoscopy for detecting dysplasia on colonoscopy, comparing results to best performing video-based dysplasia models. Finally, Aim 3 will apply computer vision quantitative histology to predict the presence of dysplasia on routine colonic biopsy, leveraging state-of-the-art histologic image segmentation methods for both enhanced pathologist annotation and modeling. Optimized dysplasia model performance will be tested and piloted in a real-world digital pathology workflow to evaluate the feasibility and performance of automated dysplasia detection in clinical practice. We expect these advancements will transform IBD dysplasia assessment by eliminating the need for cumbersome mucosal interrogation methods, improving accuracy of dysplasia detection, personalizing dysplasia surveillance and management, and providing a deployable technologic solution to elevate the quality of IBD care rendered by less-experienced clinicians.