摘要
目的比较基于多参数MRI影像组学的9种机器学习模型对校正年龄4个月~2岁婴幼儿脑瘫的预测效能。方法回顾性队列研究。纳入2013年4月—2021年9月河南中医药大学第一附属医院行MRI检查的符合条件的患儿277例,其中脑瘫89例、非脑瘫188例,按1∶1比例在非脑瘫患儿中按照门诊号随机化选取89例。最终纳入178例患儿,其中男113例、女65例,校正年龄为4个月~2岁。脑瘫组与非脑瘫组178例患儿分别按照8∶2的比例随机分入训练集(142例)和测试集(36例)。对每例患儿的T1加权像(WI)、T2WI进行图像分割和特征提取,在训练组中对提取的影像组学特征采用独立样本t检验、最小绝对收缩和选择算子(LASSO)、Z分数归一化技术进行图像特征筛选。筛选出来的影像组学特征采用逻辑回归(LRA)、决策树(DT)、随机森林(RF)、k近邻(kNN)、支持向量机(SVM)、朴素贝叶斯(NB)、梯度提升机(GBM)、轻量梯度提升机(LightGBM)、极端梯度提升(XGBoost)等9种机器学习方法构建机器学习模型。在训练集、测试集中分别采用受试者操作特征曲线(ROC曲线)评估并验证9种模型对婴幼儿脑瘫的预测效能,选择预测效能较好的模型采用五折交叉验证评估其预测效能的稳定性。结果(1)脑瘫组和非脑瘫组的基线资料比较:训练集、测试集中2组患儿智力障碍、认知障碍、缺血缺氧性脑病差异均有统计学意义(P值均<0.05),校正年龄、性别、癫痫、语言障碍、视力异常、高胆红素血症、低血糖、发育畸形、新生儿肺炎、先天性心脏病差异均无统计学意义(P值均>0.05);患儿出生时胎龄、出生体质量在训练集中2组间比较差异均有统计学意义(P值均<0.05),而在测试集中2组间比较差异均无统计学意义(P值均>0.05)。(2)MRI T1WI和T2WI特征提取后每例患者获得11190个影像组学特征,在训练集中筛选出20个特征采用9种机器学习方法构建机器学习模型。在训练集中,DT、RF、GBM、LightGBM模型的灵敏度、特异度、准确度均达到1.000,AUC均为0.995,预测效能极佳;其次是LRA、SVM、XGBoost模型,预测效能均较高,灵敏度、特异度、准确度、AUC均>0.900;kNN、NB模型的预测效能相对较低,灵敏度、特异度、准确度均<0.900。在测试集中,LRA、SVM、LightGBM模型的预测效能较高,灵敏度、特异度、准确度、AUC均>0.900,其中LRA模型的预测效能最佳(AUC=0.963);而DT、RF、kNN、NB、GBM、XGBoost模型的预测效能相对较低,灵敏度、特异度、准确度均<0.900,其中DT模型的预测效能最低(AUC=0.825)。9种机器学习模型中LRA模型在训练集、测试集预测效能均较高,AUC(95%CI)分别为0.971(0.945~0.993)、0.963(0.932~0.989),灵敏度、特异度、准确度训练集均为0.940、测试集均为0.932;对LRA模型进行五折交叉验证评估其预测效能的稳定性,ROC曲线显示在训练集、测试集中LRA模型平均AUC分别为0.971、0.963,预测效能优异且稳定。结论基于多参数MRI影像组学的9种机器学习模型对4个月~2岁婴幼儿脑瘫均具有良好的预测效能,其中LRA模型的预测效能最佳且稳定性好。
Objective:This study aimed to compare the predictive efficacy of nine machine learning models based on multi-parameter MRI radiomics for cerebral palsy in infants aged 4 months to 2 years.Methods:A retrospective cohort study was performed.A total of 277 eligible children who underwent MRI examination in the First Affiliated Hospital of Henan University of Traditional Chinese Medicine from April 2013 to September 2021 were included,of whom 89 were cerebral palsy and 188 were non-cerebral palsy.In non-cerebral palsy children,89 cases were randomly selected according to the outpatient number,and the ratio was 1∶1 with cerebral palsy children.A total of 178 children were included,including 113 males and 65 females.The age adjusted ranged from 4 months to 2 years.Among the 178 children,the cerebral palsy group and the non-cerebral palsy group were randomly divided into the training set(142 cases)and the test set(36 cases)according to the ratio of 8∶2.The T 1-weighted images(T 1WI)and T 2WI of each child were segmented,and the features were extracted.In the training group,the extracted radiomics features were screened by using an independent sample t-test,least absolute shrinkage and selection operator,and Z-score normalization technology.The selected radiomics features were used to construct machine learning models using nine machine learning methods,including logistic regression(LRA),decision tree(DT),random forest(RF),k-nearest neighbor(kNN),support vector machine(SVM),naive Bayes(NB),gradient boosting machine(GBM),lightweight gradient boosting machine(LightGBM),and extreme gradient boosting(XGBoost).The receiver operating characteristic(ROC)curve was used to evaluate and verify the predictive efficacy of the nine models for cerebral palsy in the training set and test set,and the model with better predictive efficacy was selected to evaluate the stability of its predictive efficacy using fivefold cross-validation.Results:(1)By comparing the clinical baseline data between the cerebral palsy group and the non-cerebral palsy group,statistically significant differences in intellectual disability,cognitive impairment,and hypoxic-ischemic brain disease were found between the two groups in the training set and the test set(all P values<0.05).However,no statistically significant differences in age,gender,epilepsy,language disorder,visual impairment,hyperbilirubinemia,hypoglycemia,developmental malformation,neonatal pneumonia,and congenital heart disease were observed(all P values>0.05).In addition,statistically significant differences in gestational age and birth weight were found between the two groups in the training set(all P values<0.05),but no statistically significant differences were found between the two groups in the test set(all P values>0.05).(2)After MRI and T 1WI and T 2WI feature extraction,11190 radiomic features were obtained for each patient.A total of 20 features were selected in the training set,and a machine learning model was constructed using nine machine learning methods.In the training set,the sensitivity,specificity,and accuracy of the DT,RF,GBM,and LightGBM models all reached 1.000,and AUC was greater than 0.990,indicating the best prediction performance,followed by the LRA,SVM,and XGBoost models,which all had high prediction performance,with sensitivity,specificity,accuracy,and AUC all greater than 0.900.The prediction performance of the kNN and NB models was relatively low,with sensitivity,specificity,and accuracy all less than 0.900.In the test set,the prediction performance of the LRA,SVM,and LightGBM models was high,with sensitivity,specificity,accuracy,and AUC all greater than 0.900,among which the LRA model had the best prediction performance(AUC=0.963).On the contrary,the prediction performance of the DT,RF,kNN,NB,GBM,and XGBoost models was relatively low,with sensitivity,specificity,and accuracy all less than 0.900,among which the DT model had the lowest prediction performance(AUC=0.825).Among the nine machine learning models,the LRA model had high predictive efficacy in the training set and the test set,with AUC(95%CI)of 0.971(0.945-0.993)and 0.963(0.932-0.989),respectively.The sensitivity,specificity,and accuracy of the training set were all 0.940,and those of the test set were all 0.932.The stability of the predictive efficacy of the LRA model was evaluated by fivefold cross-validation,and the ROC curve showed that the average AUC of the LRA model in the training set and test set was 0.971 and 0.963,respectively,indicating excellent and stable prediction efficiency.Conclusion:The nine machine learning models based on multi-parameter MRI imaging have good predictive efficacy for cerebral palsy in infants aged 4 months to 2 years,among which the LRA model has the best predictive efficacy and good stability.
作者
贾真
黄婷婷
卞益同
刘军
杨健
李贤军
Jia Zhen;Huang Tingting;Bian Yitong;Liu Jun;Yang Jian;Li Xianjun(Department of Medical Imaging,the First Affiliated Hospital of Xi'an Jiaotong University/Shaanxi Engineering Research Center of Computational Imaging and Medical Intelligence/Xi'an Key Laboratory of Medical Computational Imaging,Xi'an 710061,China;School of Future Technology,Xi'an Jiaotong University,Xi'an 710049,China;Department of Radiology,First Affiliated Hospital of Henan University of Traditional ChineseMedicine,Zhengzhou 450000,China)
出处
《中华解剖与临床杂志》
2025年第1期1-8,共8页
Chinese Journal of Anatomy and Clinics
基金
国家自然科学基金(82272618, 82204933)
陕西省重点研发计划(2024SF-ZDCYL-01-01)
西安交通大学第一附属医院重大新医疗新技术项目(XJYFY-2021ZD05)
西安交通大学第一附属医院国家医学中心建设揭榜项目(2023GYZX04)
西安交通大学第一附属医院临床研究中心项目(XJTU1AF-CRF-2020-005)。
关键词
脑性瘫痪
磁共振成像
影像组学
机器学习
预测
儿童
Cerebral palsy
Magnetic resonance imaging
Radiomics
Machine learning
Prediction
Child
