Dosiomics model for predicting radiation-induced temporal lobe injury in nasopharyngeal carcinoma after intensity-modulated radiotherapy
10.3760/cma.j.cn113030-20240710-00267
- VernacularTitle:基于剂量组学模型预测鼻咽癌调强放疗后放射性颞叶损伤
- Author:
Junyi LIU
1
;
Yang LI
;
Li WANG
;
Jiawei ZHOU
;
Ting QIU
;
Han GAO
;
Yinsu ZHU
;
Guanyu YANG
;
Shengfu HUANG
;
Xia HE
;
Lirong WU
Author Information
1. 江苏省肿瘤医院/江苏省肿瘤防治研究所/南京医科大学附属肿瘤医院放疗科，南京 210009
- Publication Type:Journal Article
- Keywords:
Nasopharyngeal carcinoma;
Dosiomics;
Intensity-modulated radiotherapy;
Machine learning;
Radiation-induced temporal lobe injury;
Prediction model
- From:
Chinese Journal of Radiation Oncology
2025;34(3):240-248
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To investigate and validate the performance of a dosiomics model that utilized 3D dose distribution to forecast radiation-induced temporal lobe injury (RTLI) in nasopharyngeal carcinoma (NPC) patients following intensity-modulated radiotherapy (IMRT).Methods:Clinical data of 3578 patients diagnosed with NPC admitted to Jiangsu Cancer Hospital from January 2011 to December 2021 were retrospectively analyzed. According to the inclusion and exclusion criteria, 97 NPC patients who developed RTLI were assigned into the case group. A 1:1 propensity score matching (PSM) method was used to match 97 NPC patients without RTLI as the control group. Patients were assigned into the training cohort ( n=135) and the validation cohort ( n=59) at a 7:3 ratio by simple random method. Dosiomics features were extracted from the patients' three-dimensional dose distribution maps. Spearman rho and the least absolute shrinkage and selection operator regression were used to select dosiomics features. Clinical features were collected and screened by univariate and multivariate analyses. Eight machine learning classifiers were then trained to build dosiomics models and clinical models, respectively. The area under the ROC curve (AUC), sensitivity, and specificity were calculated to compare the predictive performance of the dosiomics and clinical models. Multivariate analysis was conducted using logistic regression to assess the influencing factors, while comparisons of the ROC curves between two different models were performed using the DeLong test. Results:A total of 1130 dosiomics features were extracted from the three-dimensional dose distribution maps, and 14 features were retained for model building after feature selection. The model based on the support vector machine (SVM) classifier achieved the highest AUC value of 0.977 (95% CI: 0.949-1.000) in the validation cohort, with an AUC of 1.000 (95% CI: 1.000-1.000) in the training cohort. By conducting univariate and multivariate analyses of the patients' clinical features, 2 clinical features were retained to build the clinical model. The model based on the SVM classifier achieved the optimal AUC value of 0.667 (95% CI: 0.523-0.810) in the validation cohort, with an AUC of 0.804 (95% CI: 0.730-0.878) in the training cohort. DeLong test showed that the difference between the dosiomics and clinical models was statistically significant ( P<0.05). Conclusion:The dosiomics model based on 3D dose distribution yields high predictive performance for RTLI in NPC patients after IMRT, which surpasses the clinical feature model, providing a new approach for early clinical prediction of RTLI.
