Development and Clinical Validation of an Automated Artificial Intelligence System for Bowel Wall Thickness Assessment in Adult Inflammatory Bowel Diseases.
Authors
Affiliations (3)
Affiliations (3)
- Luigi Sacco University Hospital, Gastroenterology Unit, Department of Biomedical and Clinical Sciences, "L.Sacco" University Hospital, Lombardy, Italy, Milan.
- IRCCS Ospedale Casa Sollievo della Sofferenza, Gastroenterology, Italy, San Giovanni Rotondo.
- Samsung Group, 3Clinical Research Group, Samsung Healthcare, Seoul, Korea (the Republic of), Seoul.
Abstract
Intestinal ultrasound (IUS) is used to assess and monitor inflammatory bowel disease (IBD). Bowel wall thickness (BWT) is its key marker, but accurate measurement is operator-dependent and requires expertise. Artificial intelligence (AI) may reduce variability and support IUS implementation. We developed and validated an AI model to detect bowel wall layers and measure BWT. We developed a deep learning system (BowelAssist,BA) based on pre-trained convolutional neural network. A dataset of 9000 images (711 patients, 421 IBD) was used for training (80%), validation (10%) and testing (10%). Ground-truth annotations of bowel wall layers for 3000 segments were provided by an expert sonographer. External validation was performed live in 35 patients (67 bowel segments), with additional blinded evaluation by a trained operator in 19 (30 segments). BA performance -including BWT measurement, wall-layer recognition, and intra- and inter-observer agreement- was compared with manual measurements using Pearson's correlation and Bland-Altman analysis. BA analysis was feasible in 94.7% of measurements. It showed excellent agreement with manual BWT measurements (r=0.88 for cross-sectional and r=0.77 for longitudinal scans), and low mean absolute errors (<0.45mm). Sensitivity and specificity for detecting pathological BWT were 83.6% and 95.7%. Intra-observer reproducibility of BA measurements was >0.90 and inter-observer reproducibility between expert and trained operators was 0.53 for cross-sectional and 0.57 for longitudinal scans. Stratification recognition showed sensitivity of 90.1% and specificity of 83.8%. AI-based model provides accurate, reproducible assessment of BWT supporting its integration into clinical and research workflows for objective evaluation of IBD activity.