1335. Combining Radiological sMARIA Evaluation and MRE Radiomics to Risk Stratify Patients Needing Early Surgical Intervention in Crohn’s Disease
Authors * Denotes Presenting Author
  1. Kaustav Bera *; University Hospitals Cleveland Medical Center
  2. Prathyush Chirra; Case Western Reserve University
  3. Anamay Sharma; University Hospitals Cleveland Medical Center
  4. Jeffry Katz; University Hospitals Cleveland Medical Center
  5. Maneesh Dave; University of California Davis
  6. Raj Paspulati; H. Lee Moffitt Cancer Center & Research Institute
  7. Satish Viswanath; Case Western Reserve University
Crohn’s disease (CD) is an inflammatory bowel disease affecting up to 1.3% of the US population treated via pharmacologic therapy and eventually surgery (1). Advanced non-invasive MR enterography (MRE) imaging (2) together with radiological scoring (such as the simple MRE index of activity or sMARIA) is shown to accurately diagnose active CD, but is not prognostic of disease outcomes or need for surgery (3,4). Radiomic features that quantify subtle aspects of heterogeneity on MRE have shown promise for characterizing disease activity and fibrosis in CD (5,6). In this work, we present the first effort at integrating radiologist assessment of active CD on MRE (via sMARIA) with radiomic features to more accurately predict need for surgery in CD.

Materials and Methods:
In this retrospective, IRB-approved, single-center study, 73 patients (training - 50; validation - 23) with CD and an MRE within three months of initiating pharmacologic therapy were included. An experienced GI radiologist annotated the terminal ileum (TI) on MRE and computed the sMARIA score for the TI segment. 732 radiomic features were extracted from TI for each patient and ranked for differentiating risk groups within training set based on cross-validated machine learning performance. Random forest classifier models were built off (i) sMARIA, (ii) clinical variables (disease behavior, demographic information, serum markers), (iii) radiomics, and (iv) combination of all 3 factors. Models were evaluated via area under the ROC curve (AUC) in validation set for predicting high and low risk groups based on need for surgery within one year of MRE, as well as time to surgical intervention.

Eight radiomic features capturing textural heterogeneity within TI were significantly associated (p<0.01) with risk of surgery within one year of therapy with a hold-out validation AUC of 0.74. Meanwhile, sMARIA had a validation AUC of 0.77 while clinical variables were not significantly associated with risk of surgery (AUC=0.63). Integrating sMARIA, radiomics, and clinical variables yielded significantly improved overall performance in differentiating risk groups (validation AUC =0.83, p<0.01). Kaplan-Meier analysis of integrated model yielded a hazard ratio of 4.21 (p=6.9 e-06) and concordance index of 0.71 in predicting time to surgery after MRE.

Integration of radiomic features with radiologist sMARIA scoring on MRE images and clinical variables most accurately risk stratified severe CD patients needing surgery. Identifying high-risk CD patients requiring earlier surgery via non-invasive MRE assessment and radiomics will be critical for personalizing recommendation of pharmacological management vs surgical intervention to ensure optimal patient outcomes in CD.