Application of CT radiomics analysis to predict symptomatic radiation pneumonitis for lung cancer
10.3760/cma.j.cn112271-20210730-00301
- VernacularTitle:肺癌放疗患者症状性放射性肺炎预测的CT影像组学研究
- Author:
Yan KONG
1
;
Jia WU
;
Xianding WEI
;
Xudong KONG
;
Erwen BAO
;
Zongqiong SUN
;
Jianfeng HUANG
Author Information
1. 江南大学附属医院放疗科，无锡 214122
- Keywords:
Lung cancer;
Radiotherapy;
Radiation pneumonitis;
Radiomics;
CT
- From:
Chinese Journal of Radiological Medicine and Protection
2022;42(2):115-120
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To build a predictive model for symptomatic radiation pneumonitis(RP) using the pretreatment CT radiomics features, clinical and dosimetric data of lung cancer patients by using machine learning method.Methods:A retrospective analysis of 103 lung cancer patients who underwent radiotherapy in the Affiliated Hospital of Jiangnan University from November 2018 to April 2020 was performed. Total normal lung tissues were segmented as an interested volume in pretreatment CT images, and then 250 radiomics features were extracted. The correlations of RP and clinical or dosimetric features were firstly investigated with univariate analysis. Then all clinical data, dosimetric data and CT radiomics features were collected and considered as predictors for modeling of RP grade ≥ 2. Features were selected through LASSO machine learning method, and the predictive model was built. Finally, nomogram for risk of RP were obtained according to the selected features.Results:The result of univariate analysis showed that symptomatic RP was significantly correlated with lung dosimetric parameters including mean lung dose (MLD), V20 Gy and V30 Gy( t=2.20, 2.34 and 2.93, P<0.05). Four features, including lung dose volume percentage V30 Gyand three radiomics features, entropy feature of GLCM, mean and median feature of wavelet histogram were selected among all clinical, dosimetric features and radiomics features. AUC of the predicted model obtained from selected features reached 0.757. For convenient clinical use, the nomogram were obtained, and then personalized RP risk prediction and early intervention could be performed according to this nomogram. Conclusions:Pretreatment CT radiomics and dosimetric features can be used in predicting symptomatic RP, which will be useful for advanced intervention treatment.
