Abstract Background Intestinal ultrasound (IUS) is increasingly used for the non-invasive assessment and monitoring of inflammatory bowel disease (IBD). Bowel wall thickness (BWT) is the key sonographic marker of inflammation. We developed and validated an artificial intelligence (AI) model capable of automatically detecting and measuring bowel wall layers and thickness. Methods In this prospective study, we developed a deep learning system (BowelAssist, BA) based on a pre-trained convolutional neural network. A dataset of 9000 IUS images and 3000 videos from 711 patients (290 non-IBD, 421 IBD) was used for model training (80%), validation (10%), and testing (10%). Ground-truth annotations of bowel wall interfaces and layers for 1500 cross-sectional and 1500 longitudinal segments were provided by an expert sonographer (50,000 IUS exams). External validation was performed live in 35 patients by the expert operator, who evaluated 66 cross-sectional and 64 longitudinal segments, and in a subset of 22 patients also by an operator in advanced training. Video clips of the same segments were blindly reviewed two weeks later. The performance of BA—including BWT measurement, assessment of bowel wall stratification, and intra- and inter-observer agreement—was compared with manual measurements using Pearson’s correlation, intraclass correlation coefficients (ICC), and Bland–Altman analysis. Results In the external validation cohort, BA analysis was feasible for 494 measurements, with 26 (5.2%) unfeasible assessments. Of the feasible assessments, 25 were classified as poor quality and 443 as high quality. In all feasible measurements, BA achieved excellent agreement with manual BWT measurements in both cross-sectional and longitudinal scans (expert: r = 0.94, ICC = 0.97; trained operator: r = 0.93, ICC = 0.96). The mean absolute error compared with manual measurements was 0.29 mm for the expert and 0.39 mm for the trained operator. Sensitivity and specificity of Bowel Assist for detecting pathological thickening (BWT≥3 mm) were 90.0% and 96.5%, respectively, in high-quality assessments. Intra-observer reproducibility—comparing live assessments with blinded reviews of video clips of the same segments — was high (r = 0.94–0.93). Inter-observer agreement between the expert and the trained operator, based on blinded BA measurements performed live, was substantial (ICC=0.96). Automated recognition of bowel wall stratification exceeded 90% accuracy. Conclusion Our AI-based model accurately measures BWT and stratification on IUS, achieving reproducibility comparable to expert manual assessment. These findings support the integration of AI-assisted IUS into clinical and research workflows for the objective evaluation of disease activity and transmural healing in IBD. Conflict of interest: Prof. Dr. Maconi, Giovanni: Personal Fees: Abbvie, Alfa-Magnoni, Janssen Cilag, Samsung, Takeda Cassella, Davide Giuseppe: No conflict of interest Natalello, Gabriele: No conflict of interest Marra, Antonella: No conflict of interest Pradeep, Kumar Anand: Samsung
Maconi et al. (Thu,) studied this question.