基于颅脑CT影像组学模型初筛颞骨骨折的初步探讨  

Preliminary study of temporal bone fracture based on brain CT radiomics model

作　　者：王永芹[1] 吕赛群[1] 陈志凡[1] 周春红 杨永宏 单旭征[1] 彭涛[1] WANG Yongqin;Lü Saiqun;CHEN Zhifan;ZHOU Chunhong;YANG Yonghong;SHAN Xuzheng;PENG Tao(Department of Radiology,the Affiliated Hospital of Chengdu University,Chengdu 610081,China)

机构地区：[1]成都大学附属医院放射科,四川成都610081

出　　处：《实用放射学杂志》2023年第8期1232-1235,共4页Journal of Practical Radiology

基　　金：成都市卫生健康委员会科研课题项目(2021045,2020177);成都市金牛区医学会科研课题重点项目(JNKY2021-12)。

摘　　要：目的探讨颅脑CT影像组学模型辅助初步筛查颞骨骨折的可行性。方法纳入外伤后行头部CT平扫符合标准的患者95例(存在颞骨骨折的患者41例,右侧20例,左侧21例,均为单侧颞骨骨折,两侧颞骨均无骨折的患者54例),手动分割颞骨并提取其影像组学特征,用最小绝对收缩和选择算子(LASSO)方法进行特征选择,以5折交叉验证分组,建立逻辑回归(LR)、二次判别分析(QDA)、支持向量机(SVM)3种颞骨训练组模型,并进行验证。绘制受试者工作特征(ROC)曲线,得出各模型的曲线下面积(AUC),DeLong检验比较模型间AUC有无差异。结果5折交叉验证得出训练组和验证组LR、QDA、SVM的AUC分别为0.991、0.966、0.992和0.972、0.947、0.977。训练组QDA模型AUC分别与LR、SVM比较,P<0.05,LR和SVM模型AUC比较,P>0.05;验证组LR、QDA、SVM的AUC比较,P>0.05。结论基于颅脑CT影像组学模型具有较好初步筛查颞骨骨折的能力,LR、SVM2种模型效能近似,且优于QDA模型。Objective To investigate the feasibility of brain CT radiomics model in assisted preliminary screening for temporal bone fractures.Methods A total of 95 patients(41 patients with unilateral temporal bone fractures,20 patients on the right side and other 21 patients on the left side respectively,and 54 patients without temporal bone fractures)who met the criteria for head CT scan after trauma were included.Temporal bones were manually segmented and their radiomics features were extracted,and features selected by least absolute shrinkage and selection operator(LASSO),and grouped by five-fold cross validation.Three temporal bone training grouped models with logistic regression(LR),quadratic discriminant analysis(QDA)and support vector machine(SVM)were established respectively and verified.Receiver operating characteristic(ROC)curves plotting was used to obtain the area under the curve(AUC)of each model.DeLong test was performed to compare the difference of AUC between models.Results The AUC of five-fold cross validation were 0.991,0.966,0.992 and 0.972,0.947,0.977 for LR,QDA and SVM in the training and validation groups respectively.The AUC of QDA model in the training group compared with LR and SVM respectively,P<0.05,and the AUC of LR and SVM models compared,P>0.05.The AUC of LR,QDA and SVM in the validation group compared,P>0.05.Conclusion There is a good ability to screen temporal bone fracture initially based on the brain CT radiomics model,the LR model has similar efficacy to SVM model and is superior to QDA model.

关 键 词：影像组学 颞骨 骨折 机器学习 纹理分析 

分 类 号：R445[医药卫生—影像医学与核医学] R683.5[医药卫生—诊断学] TP181[医药卫生—临床医学]