Computed tomography-based radiomics nomogram using machine learning for predicting 1-year surgical risk after diagnosis of Crohn's disease.

BACKGROUND

Identifying patients with aggressive Crohn's disease (CD) threatened by a high risk of early onset surgery is challenging.

PURPOSE

We aimed to establish and validate a radiomics nomogram to predict 1-year surgical risk after the diagnosis of CD, thereby facilitating therapeutic strategies making.

METHODS

Patients with CD who had undergone baseline computed tomography enterography (CTE) examination at diagnosis were recruited and randomly divided into training and test cohorts at a ratio of 7:3. Enteric phase CTE images were obtained. Inflamed segments and mesenteric fat were semiautomatically segmented, followed by feature selection and signature building. A nomogram of radiomics was constructed and validated using a multivariate logistic regression algorithm.

RESULTS

A total of 268 eligible patients were retrospectively included, 69 of whom underwent surgery 1-year after diagnosis. A total of 1218 features from inflamed segments and 1218 features from peripheral mesenteric fat were extracted, and reduced to 10 and 15 potential predictors, respectively, to construct two radiomic signatures. By incorporating the radiomics signatures and clinical factors, the radiomics-clinical nomogram showed favorable calibration and discrimination in the training cohort, with an area under the curve (AUC) of 0.957, which was confirmed in the test set (AUC, 0.898). Decision curve analysis and net reclassification improvement index demonstrated the clinical usefulness of the nomogram.

CONCLUSIONS

We successfully established and validated a CTE-based radiomic nomogram with both inflamed segment and mesenteric fat simultaneously evaluated to predict 1-year surgical risk in CD patients, which assisted in clinical decision-making and individualized management.