The value of machine learning models based radiomics for predicting high-risk molecular subtypes of lower-grade gliomas
10.3760/cma.j.cn112149-20250609-00330
- VernacularTitle:基于影像组学的机器学习模型预测较低级别胶质瘤高危分子亚型的价值
- Author:
Xiangli YANG
1
;
Guoqiang YANG
;
Wenju NIU
;
Xueting LI
;
Yan TAN
;
Xiaochun WANG
;
Lizhi XIE
;
Hui ZHANG
Author Information
1. 山西医科大学第一医院影像科，太原030001
- Publication Type:Journal Article
- Keywords:
Gliomas;
Lower-grade;
Magnetic resonance imaging;
Radiomics;
High-risk molecular subtypes
- From:
Chinese Journal of Radiology
2025;59(8):909-916
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To evaluate the clinical utility of machine learning model based radiomics in predicting high-risk molecular subtypes of lower-grade gliomas(LrGGs).Methods:This was a cross-sectional study. A total of 287 patients diagnosed with LrGGs in the First Hospital of Shanxi Medical University, Shanxi Provincial People′s Hospital, and the Third Hospital of Shanxi Medical University from January 2011 to September 2023 were retrospectively collected, including 166 males and 121 females; 114 cases of high-risk molecular subtypes and 173 cases of non-high-risk molecular subtypes. All patients were divided into 201 cases in the training set and 86 cases in the test set according to 7∶3 in simple randomized grouping method. All patients underwent contrast-enhanced T 1WI (CE-T 1WI) and T 2-weighted fluid-attenuated inversion recovery sequence imaging (T 2-FLAIR), and the imaging features of high-risk and non-high-risk molecular subtypes were analyzed. Analysis of variance, recursive feature elimination, and Kruskal-Wallis were used for radiomics feature screening, and a support vector machine (SVM) classifier was used to construct a radiomics-based classifier model. Univariate and multivariate logistic regression were used to analyze clinical variables independently influencing high-risk molecular subtypes of LrGGs to construct a clinical model; a combined model was developed by integrating radiomics labels and clinical variables. Receiver operating characteristic curve and area under the curve (AUC), calibration curve, and decision curve were used to compare the predictive performance of different models. Results:The patient′s age ( OR=1.042, 95% CI 1.018-1.068, P=0.001), pathological grade ( OR=2.270, 95% CI 1.212-4.311, P=0.011), MGMT methylation status ( OR=0.456, 95% CI 0.238-0.866, P=0.017), and ependymal involvement ( OR=7.335, 95% CI 2.929-18.370, P<0.001) were independent influencing factors for the high-risk molecular subtype of LrGGs, and a clinical model was developed based on these factors. An SVM model was constructed based on 12 radiomics features (3 radiomics features based on CE-T 1WI and 9 radiomics features based on T 2-FLAIR). The radiomics score of the probability output by the SVM model was combined with age, pathological grade, MGMT methylation status, and ependymal involvement to develop a combined model. The AUC values of the SVM model for predicting the high-risk molecular subtype of LrGGs were 0.824 and 0.859 in the training set and test set, respectively; the AUC values of the clinical model in the training set and test set were 0.759 and 0.721, respectively; and the AUC values of the combined model in the training set and test set were 0.823 and 0.815, respectively. The combined model had a high clinical net benefit. Conclusion:The machine learning MRI radiomics model can preoperatively predict high risk molecular subtypes of LGGrs, assist in individualized treatment decisions.
