Computed tomography enterography-based deep learning radiomics models of intestinal lesions and perienteric fat for predicting Crohn's disease activity: a multicenter cohort study.

BACKGROUND

Accurate assessment of Crohn's disease (CD) activity is critical for improving patient prognosis, while current imaging methods lack quantitative objectivity. Therefore, this study aimed to develop an integrated model based on computed tomography enterography (CTE) using radiomics, deep learning, and clinical characteristics for noninvasively predicting CD activity.

METHODS

A total of 259 CD patients were retrospectively enrolled from three medical centers. With patients from one center serving as an independent external validation cohort (n=68), patients from the other two centers were randomly divided into a training (n=133) and a testing (n=58) cohort in a 7:3 ratio. Independent predictors were selected from clinical features to construct a clinical model. Based on the imaging features of intestinal lesions and perienteric adipose tissue, radiomics models were constructed using least absolute shrinkage and selection operator (LASSO) regression, and deep learning radiomics models were developed using features extracted by ResNet-18 and ResNet-34 architectures. Finally, an integrated model combining the radiomics, deep learning, and independent clinical predictors was constructed. The area under the receiver operating characteristic curve (AUC) was used to assess model performance.

RESULTS

The combined model demonstrated optimal performance, with an AUC of 0.984 [95% confidence interval (CI): 0.970-0.999], 0.902 (95% CI: 0.804-1.000) and 0.884 (95% CI: 0.809-0.960) in the training, testing and validation cohorts, respectively. Among other models, the deep learning radiomics model for intestinal lesions achieved the highest AUC of 0.938 (95% CI: 0.901-0.976), 0.865 (95% CI: 0.767-0.963) and 0.841 (95% CI: 0.745-0.938) in the training, testing and validation cohorts. No statistically significant difference was observed between the intestinal lesion-based radiomics model and the perienteric fat-based radiomics model.

CONCLUSIONS

The combined model integrating the radiomics based on intestinal lesions and perienteric fat, deep learning, and clinical characteristics demonstrated the best performance in predicting CD activity and can provide novel insights for clinical decision-making.