Machine Learning Diagnosis of Small Bowel Crohn Disease Using T2-Weighted MRI Radiomic and Clinical Data.

Background: Radiologists show variable diagnostic performance and considerable inter-reader variability when interpreting MR enterography (MRE) examinations for suspected Crohn disease (CD). Objective: To develop a machine learning-based method for predicting ileal CD using radiomic features of ileal wall and mesenteric fat from noncontrast T2-weighted MR images, and to compare performance with that of expert radiologists. Methods: This single-institution study included retrospectively identified patients who underwent MRE for suspected ileal CD from January 1, 2020 to January 31, 2021, and prospectively enrolled participants (with newly diagnosed ileal CD or serving as healthy controls) from December 2018 to October 2021. Using axial T2-weighted SSFSE images, a radiologist selected two slices show-ing greatest terminal ileal wall thickening; four ROIs were segmented, and radiomic features were extracted from each ROI. Following feature selection, support-vector machine models were trained to classify presence of ileal CD. Three fellowship-trained pediatric abdominal radiologists inde-pendently classified presence of ileal CD on SSFSE images. Reference standard was clinical diag-nosis of ileal CD following positive endoscopy and biopsy. Radiomic-only, clinical-only, and radi-omic-clinical ensemble models were trained and evaluated using nested cross-validation. Results: The study included 135 patients (mean age, 15.2±3.2 years; 67 female, 68 male); 70 were diagnosed with ileal CD. The three radiologists had accuracies of 83.7% (113/135), 86.7% (117/135), and 88.1% (119/135) for diagnosing CD; consensus accuracy was 88.1%. Inter-radiologist agreement was substantial (kappa=0.78). Best-performing ROI was bowel-core (AUC=0.95, accuracy=89.6%); other ROIs had worse performance (whole-bowel AUC=0.86; fat-core AUC=0.70; whole-fat AUC=0.73). For clinical-only model, AUC was 0.85 and accuracy was 80.0%. Ensemble model combining bowel-core radiomic and clinical models achieved AUC of 0.98 and accuracy of 93.5%. Bowel-core radiomic-only model demonstrated better accuracy than radiologist 1 (P=.009) and radiologist 2 (P=.02), but not radiologist 3 (P>.99) or radiologists' consensus (P=.05). Ensemble model demonstrated better accuracy than radiologists' consensus (P=.02). Conclusions: A radiomic machine learning model predicted CD diagnosis with better performance than two of three expert radiologists. Model performance improved when ensembled with clinical data. Clinical Impact: Deployment of a radiomic-based model using T2-weighted MR data could de-crease inter-radiologist variability and increase diagnostic accuracy for pediatric CD.