摘要
目的通过时间验证评价基于多参数磁共振成像(mp-MRI)、纹理分析与临床资料所建立的复合机器学习模型对临床显著性前列腺癌(csPCa)的诊断效能和稳定性。方法回顾性分析246例前列腺mp-MRI的患者资料,将mp-MRI相关特征(包括PI-RADS V2.1评分、动态增强定量参数)、纹理分析和部分临床参数进行组合,进行降维和特征选择,建立SVM和Logistic模型,并进行内部验证和时间验证,分别运用受试者工作特征曲线、决策曲线分析(DCA)比较两模型的诊断效能及临床获益。结果优选出的关键参数有:T_(2).Quantile10、T_(2).Quantile95、ADC.MinIntensity、ADC.Uniformity、ADC.Quantile75、ImageStd、VeStd、Ve0.1、Vp0.75、TTPMax、DWI(PI-RADS)、PIRADS、age、tPSA,Logistic模型中T_(2).Quantile95为csPCa独立预测参数(P<0.05)。SVM和Logistic模型对内部验证组进行分类的曲线下面积(AUC)分别为0.96、0.88,对时间验证组进行分类的AUC分别为0.80和0.77,两种模型的AUC之间的差异无统计学意义(P>0.05);内部验证组与时间验证组两模型的AUC之间的差异有统计学意义(P<0.05)。DCA表明:概率阈值为0.12~0.49时,Logistic模型净获益高于SVM模型;概率阈值为0.50~0.95时,SVM模型净获益高于Logistic模型。结论14个关键参数中T_(2).Quantile95为独立预测参数,基于这些参数建立的Logistic与SVM模型对csPCa都能取得较好的诊断效能,两模型的稳定性较好,概率阈值<0.5时,Logistic模型的净获益较大,概率阈值>0.5时,SVM模型的净获益较大。
Objective Temporal validation is used to evaluate the diagnostic efficacy and stability of a composite machine learning model based on multi-parameter magnetic resonance(mp-MRI),texture analysis and clinical data for clinically significant prostate cancer(csPCa).Methods The patients underdone magnetic resonance imaging were analyzed retrospectively,and mp-MRI related features(including PI-RADS v2.1 score,dynamic enhancement quantitative parameters),texture analysis and some clinical Parameters are combined,and then perform dimensionality reduction and feature selection to establish SVM and Logistic models,and perform internal validation and temporal validation.Receiver operating characteristic curve(ROC)and decision curve analysis(DCA)are used to compare the diagnostic efficacy and clinical benefits of the two models.Results The optimized key parameters are:T_(2).Quantile10,T_(2).Quantile95,ADC.MinIntensity,ADC.Uniformity,ADC.Quantile75,mageStd,VeStd,Ve0.1,Vp0.75,TTPMax,DWI(PI-RADS),PIRADS,age、tPSA,of which T_(2).quantile 95 is the independent prediction parameter of csPCa(P<0.05).The areas under the curve of SVM andLogistic model for the classification of internal validation group were 0.96 and 0.88 respectively,and the AUC of temporal validation group were 0.80 and 0.77 respectively.There was no significant difference between the AUC of the two models;There was significant difference in AUC between internal validation group and temporal validation group(P<0.05).DCA shows that when the probability threshold is 0.12-0.49,the net benefit ofLogisticmodel is higher than that of SVM model;When the probability threshold is 0.50-0.95,the net benefit of SVM model is higher than that ofLogisticmodel.Conclusions Among the 14 key parameters,T_(2).quantile95 is an independent prediction parameter.TheLogisticand SVM models based on these parameters can achieve good diagnostic efficiency for csPCa.The stability of the two models is good.When the probability threshold is less than 0.5,the net benefit ofLogisticmodel is greater,and when the probability threshold is more than0.5,the net benefit of SVM model is greater.
作者
李林
伍兵
彭涛
肖建明
吕赛群
曾小辉
LI Lin;WU Bing;PENG Tao(Department of Radiology,West China Hospital,Sichuan University,Chengdu,Sichuan Province 610041,China.)
出处
《临床放射学杂志》
北大核心
2022年第11期2082-2086,共5页
Journal of Clinical Radiology
基金
成都市医学科研课题项目(编号:2021036)
成都市医学科研课题项目(编号:2020177)。
关键词
多参数磁共振
临床显著性前列腺癌
机器学习模型
时间验证
Multi-parameter magnetic resonance
Clinically significant prostate cancer
Machine learning model
Temporal validation