Artificial Intelligence-Based Localization of Small Bowel Anatomical Transition Zones in Crohn’s Disease Using Capsule Endoscopy

Background: Small bowel (SB) capsule endoscopy (SBCE) is a key tool for Crohn’s disease (CD), but interpretation is time-consuming. Artificial intelligence (AI) could assist, yet automated SB segmentation, prerequisite for analysis, may be challenging in CD patients. Aims: To evaluate an AI model detecting the pylorus and ileocolonic junction in CD SBCEs. Methods: SBCE videos from 60 CD patients were annotated by three readers. A ResNet-18 convolutional neural network classified frames into stomach, SB, or colon, with Viterbi-based anatomically constrained segmentation, using five-fold cross-validation. Accuracy was assessed by discrepancies relative to expert annotations. Univariate analysis explored associated factors. Expert review was conducted for poorly localized cases. Results: The model was trained on 596,980 images, including 426,674 SB images. Median localization error was 2 frames (1 second) for the pylorus and 118.5 frames (~5 minutes) for the ileocolonic junction. In 26 videos (43.3%), the model localized both junctions with a temporal discrepancy < 1 minute compared with experts. In 53.3% of cases, AI detection of the ileocolonic junction preceded expert annotation by a median of 46 minutes, potentially leading to missed terminal ileal lesions. No factors were significantly associated with localization accuracy. Expert review highlighted inadequate cleanliness, capsule stagnation and back-and-forth movements as main contributors to poor AI performance. Conclusions: Our model successfully localizes both SB junctions in nearly half of SBCE videos from CD patients. It consistently identifies the pylorus but shows modest performance at the ileocolonic junction. Capsule dynamics and bowel cleanliness remain major challenges for automated SB segmentation.