A multicenter study of brain T 2WI lesions radiomics machine learning models distinguishing multiple sclerosis and neuromyelitis optica spectrum disorder
10.3760/cma.j.cn112149-20220106-00013
- VernacularTitle:基于脑部T 2WI病灶影像组学的机器学习模型鉴别多发性硬化和视神经脊髓炎谱系疾病的多中心研究
- Author:
Ting HE
1
;
Yi MAO
;
Zhi ZHANG
;
Zhizheng ZHUO
;
Yunyun DUAN
;
Lin WU
;
Yuxin LI
;
Ningnannan ZHANG
;
Xuemei HAN
;
Yanyan ZHU
;
Yao WANG
;
Xiao LIANG
;
Yongmei LI
;
Haiqing LI
;
Fuqing ZHOU
;
Ya′ou LIU
Author Information
1. 南昌大学第一附属医院影像科，南昌 330006
- Keywords:
Multiple sclerosis;
Magnetic resonance imaging;
Neuromyelitis optica spectrum disorder;
Differential diagnosis;
Radiomics
- From:
Chinese Journal of Radiology
2022;56(12):1332-1338
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To investigate the efficacy of a machine learning model based on radiomics of brain lesions on T 2WI in differentiating multiple sclerosis (MS) from neuromyelitis optica spectrum disorders (NMOSD). Methods:Totally 223 MS and NMOSD patients who were treated from January 2009 to September 2018 in Beijing Tiantan Hospital Affiliated to Capital Medical University, Donghu Branch of the First Affiliated Hospital of Nanchang University, Tianjin Medical University General Hospital, and Xuanwu Hospital of Capital Medical University were analyzed retrospectively, and according to the proportion of 7∶3, 223 patients were completely randomly divided into training set (156 cases) and test set (67 cases). A total of 74 patients with MS and NMOSD who were treated in Huashan Hospital Affiliated to Fudan University and China-Japan Friendship Hospital of Jilin University from January 2009 to September 2018 and in Xianghu Branch of the First Affiliated Hospital of Nanchang University from March 2020 to September 2021 were collected as an independent external validation set. All patients underwent brain cross-sectional MR T 2WI, radiomics features were extracted from T 2WI, and features were selected by max-relevance and min-redundancy and least absolute shrinkage and selection operator algorithms. Then various machine learning classifier models (logistic regression, decision tree, AdaBoost, random forest or support vector machine) were constructed to differentiate MS from NMOSD. The area under curve (AUC) of receiver operating characteristics was used to evaluate the performance of each classifier model in the training set, test set and external validation set. Results:Based on multi-center T 2WI, a total of 11 radiomics features related to the discrimination between MS and NMOSD were extracted and classifier models were constructed. Among them, the random forest model had the best efficiency in distinguishing MS from NMOSD, and its AUC values for distinguishing MS from NMOSD in the training set, test set and external validation set were 1.000, 0.944 and 0.902, with specificity of 100%, 76.9% and 86.0%, and sensitivity of 100%, 92.1% and 79.7%, respectively. Conclusion:The random forest model based on the radiomic features of T 2WI of brain lesions can effectively distinguish MS from NMOSD.
