A CT-based nomogram established for differentiating gastrointestinal heterotopic pancreas from gastrointestinal stromal tumor: compared with a machine-learning model

OBJECTIVE: To identify CT features and establish a nomogram, compared with a machine learning-based model for distinguishing gastrointestinal heterotopic pancreas (HP) from gastrointestinal stromal tumor (GIST). MATERIALS AND METHODS: This retrospective study included 148 patients with pathologically confirmed HP (n = 48) and GIST (n = 100) in the stomach or small intestine that were less than 3 cm in size. Clinical information and CT characteristics were collected. A nomogram on account of lasso regression and multivariate logistic regression, and a RandomForest (RF) model based on significant variables in univariate analyses were established. Receiver operating characteristic (ROC) curve, mean area under the curve (AUC), calibration curve and decision curve analysis (DCA) were carried out to evaluate and compare the diagnostic ability of models. RESULTS: The nomogram identified five CT features as independent predictors of HP diagnosis: age, location, LD/SD ratio, duct-like structure, and HU lesion/pancreas A. Five features were included in RF model and ranked according to their relevance to the differential diagnosis: LD/SD ratio, HU lesion/pancreas A, location, peritumoral hypodensity line and age. The nomogram and RF model yielded AUC of 0.951 (95% CI: 0.842–0.993) and 0.894 (95% CI: 0.766–0.966), respectively. The DeLong test found no statistically significant difference in diagnostic performance (p > 0.05), but DCA revealed that the nomogram surpassed the RF model in clinical usefulness. CONCLUSION: Two diagnostic prediction models based on a nomogram as well as RF method were reliable and easy-to-use for distinguishing between HP and GIST, which might also assist treatment planning.