Abstract
Inflammatory bowel disease (IBD) is a recurrent bowel disease that usually requires magnetic resonance enterography (MRE) for diagnosis and monitoring. However, recognition of bowel segments from MRE images by a radiologist is challenging and time-consuming. Deep learning-based medical image segmentation has shown the potential to reduce manual effort and provide automated tools to assist in disease management; however, it requires a large-scale fine-annotated dataset for training. To address this gap, we collected MRE data, including half-Fourier acquisition single-shot turbo spin-echo(HASTE) sequences with coronal orientation, from 114 patients with IBD, who received 1600-2000 mL of 2.5% mannitol. The bowel images per patient were contoured and annotated into ten segments (stomach, duodenum, small intestine, appendix, cecum, ascending colon, transverse colon, descending colon, sigmoid colon, and rectum), with fine pixel-level annotations labeled by experienced radiologists. Furthermore, we validated the efficiency of several state-of-the-art segmentation methods using this dataset. This study established a high-quality, publicly available whole-bowel segment MR dataset with benchmark results and laid the groundwork for AI research on IBD.