A deep learning framework for autonomous detection and classification of Crohn's disease lesions in the small bowel and colon with capsule endoscopy.

Background and study aims Small bowel ulcerations are efficiently detected with deep learning techniques, whereas the ability to diagnose Crohn's disease (CD) in the colon with it is unknown. This study examined the ability of a deep learning framework to detect CD lesions with pan-enteric capsule endoscopy (CE) and classify lesions of different severity. Patients and methods CEs from patients with suspected or known CD were included in the analysis. Two experienced gastroenterologists classified anonymized images into normal mucosa, non-ulcerated inflammation, aphthous ulceration, ulcer, or fissure/extensive ulceration. An automated framework incorporating multiple ResNet-50 architectures was trained. To improve its robustness and ability to characterize lesions, image processing methods focused on texture enhancement were employed. Results A total of 7744 images from 38 patients with CD were collected (small bowel 4972, colon 2772) of which 2748 contained at least one ulceration (small bowel 1857, colon 891). With a patient-dependent split of images for training, validation, and testing, ulcerations were diagnosed with a sensitivity, specificity, and diagnostic accuracy of 95.7 % (CI 93.4-97.4), 99.8 % (CI 99.2-100), and 98.4 % (CI 97.6-99.0), respectively. The diagnostic accuracy was 98.5 % (CI 97.5-99.2) for the small bowel and 98.1 % (CI 96.3-99.2) for the colon. Ulcerations of different severities were classified with substantial agreement (κ = 0.72). Conclusions Our proposed framework is in excellent agreement with the clinical standard, and diagnostic accuracies are equally high for the small bowel and colon. Deep learning approaches have a great potential to help clinicians detect, localize, and determine the severity of CD with pan-enteric CE.