Combined model based on enhanced CT texture features in liver metastasis prediction of high-risk gastrointestinal...

To investigate the use of the combined model based on clinical and enhanced CT texture features for predicting the liver metastasis of high-risk gastrointestinal stromal tumors (GISTs).

This retrospective study was conducted including 204 patients with pathologically confirmed high-risk GISTs from the Zhejiang Cancer Hospital from January 2015 to June 2021, and 76 cases of them were diagnosed with simultaneous liver metastasis. We randomly divided the cohort into a training cohort (n = 142) and a validation cohort (n = 62) with a ratio of 7:3. All volumes of interest (VOIs) of the high-risk GISTs were manually segmented on the portal venous phase CT images using the ITK-SNAP software. The least absolute shrinkage and selection operator (Lasso) algorithm was performed to determine the most valuable features from a total of 110 texture features extracted by the A-K software to reflect the texture information of the given VOIs. Texture-based predictive model was built from the selected texture features. Independent clinical risk factors were identified through univariate logistic analysis. Then, the texture-based model incorporated the clinical predictors to develop a combined model by multivariate logistic regression. Receiver operating characteristic curve, calibration curve, and decision curve analysis were utilized to analyze the discrimination capacity and clinical application value of the predictive models.

The nine optimal texture features were remained after the reduction of dimension using Lasso method. Another four clinical parameters (BMI, location, gastrointestinal bleeding, and CA125 level) were included in the clinical-based predictive model. Finally, with the combination of remaining texture and clinical features, a multivariate logistic regression classifier was built to predict the liver metastasis potential of high-risk GISTs. The remarkable classification performance of the combined model for the prediction of liver metastasis in the subjects with high-risk GISTs was obtained with area under curve (AUC) = 0.919, sensitivity = 83.9%, specificity = 89.7%, and accuracy = 84.9% in our validation group.

The texture-based radiomic signature derived from the portal venous phase CT images could predict liver metastasis of high-risk GISTs in a non-invasive way. Integrating additional clinical variables into the model further leads to an improvement of liver metastasis risk prediction.