Different radiomics models in predicting the malignant potential of small intestinal stromal tumors

Objectives: To explore the feasibility of different radiomics models for predicting the malignant potential of small intestinal stromal tumors (SISTs), and to select the best radiomics model. Methods: A retrospective analysis of 140 patients with SISTs was conducted. Radiomics features were extracted from CT-enhanced images. Support vector machine (SVM), Decision tree (DT), Conditional inference trees (CIT), Random Forest (RF), K-nearest neighbors (KNN), Back-propagation neural network (BPNet), and Bayes were used to construct different radiomics models. The clinical data and CT performance were selected using univariate analysis and to construct clinical model. Nomogram model was developed by combining clinical data and radiomics features. Model performances were assessed by using the area under the receiver operator characteristic (ROC) curve (AUC). The models' clinical values were assessed by decision curve analysis (DCA). Results: A total of 1132 radiomics features were extracted. Among radiomics models, SVM was better than DT, CIT, RF, KNN, BPNet, Bayes because it had the highest AUC with a significant difference (P<0.05). The AUC of the clinical model was 0.781. The AUC of the radiomics model was 0.910. The AUC of nomogram model was 0.938. Clinical models had the lowest AUC. Nomogram AUC were slightly higher than radiomics model, but the difference was not significant (P=0.48). The DCA of the nomogram model and radiomics model showed optimal clinical efficacy. Conclusions: The model constructed with SVM method was the best model for predicting the malignant potential of SISTs. Radiomics model and nomogram model showed high predictive value in predicting the malignant potential of SISTs.