# Oral contrast-enhanced ultrasonographic features and radiomics analysis to predict NIH risk stratification for gastrointestinal stromal tumors

## TL;DR

This study shows that combining ultrasound imaging with radiomics analysis improves prediction of risk levels for gastrointestinal stromal tumors.

## Contribution

The novel contribution is the development of a combined model using oral contrast-enhanced ultrasound and radiomics for predicting NIH risk stratification in GISTs.

## Key findings

- The US radiomics model outperformed the US model in predicting NIH risk stratification.
- The combined model integrating clinical, ultrasound, and radiomics features showed better diagnostic performance than individual models.
- The combined model performed similarly to endoscopic ultrasound in risk prediction.

## Abstract

To evaluate the value of oral contrast-enhanced ultrasonography and radiomics analysis in predicting the National Institutes of Health (NIH) staging of gastrointestinal stromal tumors (GISTs).

A retrospective cohort study was conducted on 204 patients presenting with GISTs in Tianjin Medical University Cancer Institute and Hospital from January 2020 to January 2023. The clinical profiles, oral contrast-enhanced ultrasonography (CEUS), and endoscopic ultrasound (EUS) imaging data were collected. 105 patients with high-risk and moderate-risk GISTs were classified into the high-risk group, while 99 patients with low-risk and very-low-risk GISTs were classified into the low-risk group. The ITK-SNAP software and Pyradiomics (version 3.0.1) package were used to extract a comprehensive set of ultrasonographic radiomics features from the segmented regions of interest (ROIs). The patient dataset was randomly divided into a training set and a validation set at a ratio of 7:3. Leveraging the XGBoost (XGB) algorithm within the Scikit-learn (Sklearn) machine-learning library, three distinct predictive models were developed: a clinical ultrasound imaging model (US model), an ultrasonographic radiomics model (US radiomics model), and a combined model integrating both clinical, ultrasound, and radiomics features. Additionally, 51 GIST patients from Tianjin Medical University General Hospital were included in the external validation analysis.

636 ultrasonic radiomics features from ROIs were successfully extracted. 6 key ultrasonic radiomics features were finally selected for subsequent model construction. In the internal validation set, the area under the curve (AUC), sensitivity, specificity, and accuracy for the US model, US radiomics model, combined model, and endoscopic ultrasound were 0.69, 0.62, 0.66, 0.64; 0.83, 0.85, 0.74, 0.79; 0.91, 0.86, 0.85, 0.85; and 0.94, 0.95, 0.85, 0.89, respectively. In the external validation set, the AUC, sensitivity, specificity, and accuracy for the US model, US radiomics model, combined model, and endoscopic ultrasound were 0.71, 0.65, 0.67, 0.66; 0.81, 0.77, 0.72, 0.74; 0.89, 0.85, 0.80, 0.83; and 0.90, 0.93, 0.86, 0.90, respectively. The Delong test showed a larger AUC in the US radiomics model compared with the US model (Z = 2.776, P < 0.01). The performance of the combined model was significantly better than that of the US model (Z = 4.822, P < 0.01) and the US radiomics model (Z = 2.200, P = 0.029). However, there was no significant difference in AUC between the combined model and the endoscopic ultrasound (Z = 1.150, P = 0.141). The superiority of the combined model was further demonstrated by the calibration curve (CC) and decision curve analysis (DCA) in both the internal and external validation sets.

This study demonstrates that the US radiomics model, based on oral contrast-enhanced ultrasonography images, is feasible for predicting the NIH risk stratification of gastrointestinal stromal tumors (GISTs). The combined model showed a better diagnostic performance.