Radiological-Clinical Machine-Learning Model for MRE-Based Mucosal Healing Assessment in Crohn`s Disease

Itai Guez ¹ Gili Focht ² Mary-Louise C. Greer ³ Ruth Cytter-Kuint ⁴ Li-tal Pratt ⁵ Denise A. Castro ⁵ Dan Turner ² Anne M. Griffiths ⁶ Moti Freiman ¹

¹Bio-medical Engineering, Technion Israel Institute of Technology, Israel
²Juliet Keidan Institute of Paediatric Gastroenterology and Nutrition, Shaare Zedek Medical Center, Israel
³Department of Medical Imaging, The Hospital for Sick Children, Canada
⁴Paediatric Radiology Unit, Shaare Zedek Medical Center, Israel
⁵Department of Diagnostic Imaging, Kingston Health Sciences Centre, Canada
⁶Gastroenterology, Hepatology and Nutrition, The Hospital for Sick Children, Canada

Purpose: Mucosal Healing (MH) is considered a major Crohn’s Disease (CD) treatment goal. Simple Endoscopic Score of Crohn`s Disease (SES-CD) is a standard
quantitative endoscopic score for MH (SES-CD<3). However, ileocolonoscopy which requires anesthesia is poorly tolerated by many Patients, especially in the pediatric population. The purpose of this study was to evaluate the capacity of a radiological-clinical machine-learning model to non-invasively predict terminal-ileum (TI) SES-CD from Magnetic-Resonance-Enetrography (MRE) data.

Materials and Methods: This is an ImageKids substudy in which 240 pediatric patients with CD underwent a baseline ileocolonoscopy scored by SES-CD, followed by an MRE examination and clinical data collection. We used a non-linear machine-learning random-forest (RF) model with multiple combinations of radiological and clinical biomarkers compared to a baseline linear model developed with the Magnetic-
Resonance-Index-of-Activity (MaRIA) score to determine the best combination of
radiological and clinical information for SES-CD prediction. We train each
algorithm version with an equal random set of patients used for derivation and
validation 100-folds. We determined the best radiological-clinical model by
comparing the various models’ capacity to discern between patients with and
without MH by means of averaged area under the curve over the different folds.

Results: A subsample of 78 patients had all relevant TI items scored by central
radiologists. The RF model consisted with mucosal biomarkers set (ulcers, wall
restricted diffusion, length) enriched with C-Reactive-Protein (CRP) and Fecal-
Calprotectin (FC) normalized by the relative involved segment length to compensate for their non-specific nature, produces the highest Area Under Curve (AUC) vs all other models. The difference in AUC compared to a baseline MaRIA-based model was statistically significant (0.85 vs 0.74, p<1e-9, De-Long’s test).

Conclusions: A non-linear machine-learning algorithm developed with a combination of mucosal MRE-based biomarkers and normalized clinical biomarkers improves
non-invasive MH prediction from MRE data.