Development of an artificial intelligence tool for detecting colorectal lesions in inflammatory bowel disease.

Background and Aims: Artificial intelligence has the potential to enhance the endoscopist's ability to detect polypoid and nonpolypoid dysplasia, which could result in decreased rates of colorectal cancer in patients with inflammatory bowel disease (IBD). We aimed to develop the first known model for computer-aided detection (CADe) of colorectal lesions in patients with IBD. Methods: A CADe model developed at our institution with colorectal lesions from patients without IBD was tested for baseline performance in a dataset of high-definition white-light endoscopy (HDWLE) images of IBD-associated colorectal lesions. Subsequently, we retrained the original CADe to build an IBD-CADe model using 1266 HDWLE still images and 426 dye-based chromoendoscopy still images depicting histologically proven IBD-associated colorectal lesions. The lesions were annotated by histopathology, size, morphology, and inflammation score surrounding the lesion. We evaluated the model's performance metrics before and after retraining. Results: The sensitivity of the original CADe was 50%, positive predictive value (PPV) was 97%, and accuracy was 64%. With the retrained IBD-CADe model, performance metrics for detecting lesions on HDWLE were as follows: sensitivity, 95.1%; specificity, 98.8%; PPV, 98.9%; negative predictive value, 94.7%; accuracy, 96.8%; and area under the curve, .85. IBD-CADe for chromoendoscopy images showed a sensitivity of 67.4%, specificity of 88.0%, PPV of 83.3%, negative predictive value of 74.3%, accuracy of 77.8%, and area under the curve of .65. Subgroup analysis showed a 93% sensitivity for detecting lesions 5 mm or smaller, 91% for lesions 6 to 10 mm, and 85% for those larger than 10 mm. IBD-CADe performed best for Paris classification types Ip, Is, and IIa. Of 9 lesions missed by IBD-CADe, most had a Mayo endoscopic subscore of 0 or 1. Conclusions: This model is the first step toward developing other artificial intelligence-based endoscopic tools to enhance dysplasia detection for patients with IBD.