Preoperatively Predicting Risk Stratification for GISTs ≤2 cm by Radiomics Model: A Dual-center Study

Small gastrointestinal stromal tumors (SGISTs, maximum diameter≤2 cm) still carry a risk of malignancy, and their preoperative evaluation remains a significant challenge. Radiomics, an emerging technique for analyzing image data, has yet to be employed to assess the risk stratification of SGISTs. To develop and validate a CT radiomics model for the preoperative prediction of risk stratification in patients with SGISTs.

This study enrolled 133 patients with SGISTs, including 97 in the low-grade group and 36 in the high-grade group. Patients were randomly assigned to a training set (n = 93) and a testing set (n = 40) at a ratio of 7:3. Radiomics features were extracted from preoperative CT images, and dimensionality reduction was performed using the LR-LASSO to identify the most predictive features for constructing the radiomics model. Clinical features were evaluated using univariate and multivariate logistic regression analyses to develop a clinical model. Subsequently, the optimal radiomics and clinical features were integrated to establish a combined model. Model performance was evaluated using ROC curve analysis, and a corresponding nomogram was generated to facilitate clinical application. The Delong test was used to compare the ROC curves, with a p-value < 0.05 considered statistically significant.

Univariable clinical analysis identified maximal tumour diameter as the only significant predictor, with the clinical model achieving an AUC of 0.641 (95% CI: 0.533–0.748). Among the radiomics signatures derived from multiphase CT (non-contrast to delayed phases), the model based on portal venous phase images demonstrated the highest discriminative ability, yielding the best AUC values in both the training set (AUC = 0.848, 95% CI: 0.764–0.931) and the testing set (AUC = 0.824, 95% CI: 0.696–0.953). The combined model, which integrated radiomics features with maximum tumour diameter, demonstrated improved performance, attaining an AUC of 0.862 (95% CI: 0.743–0.975) in the training set and 0.859 (95% CI: 0.743–0.975) in the testing set. Notably, the predictive performance of both the radiomics and combined models was significantly greater than that of the clinical model (DeLong test, P < 0.05). However, no statistically significant differences were observed between the AUC values of the radiomics and combined models. Calibration curves indicated a good fit, and the DCA demonstrated that both the radiomics model and the combined model provided greater clinical benefits.

The radiomics model demonstrated superior performance to the clinical model for the preoperative prediction of risk stratification in SGISTs. As a visualization tool, the nomogram of the combined model plays a critical role in optimizing early surgical resection decisions.

The radiomics model could serve as an effective tool for non-invasive risk stratification of SGISTs, offering clear advantages over risk stratification models based solely on conventional clinical parameters. This approach could support improved preoperative clinical decisionmaking.