摘要
目的 观察PET/MRI海马纹理特征诊断阿尔茨海默病(AD)及遗忘型轻度认知障碍(aMCI)的价值.方法 回顾性分析55例AD(AD组)、60例aMCI患者(aMCI组)及55名健康受试者(HC组),按7:3比例随机分为训练集与测试集,行一体化PET/MRI,获取3D T1WI和18 F-FDG PET图;对训练集提取双侧海马ROI纹理特征,分别以逻辑回归(LR)、支持向量机(SVM)及随机森林(RF)算法建立3DT1WI模型、18F-FDG PET模型及联合模型,以受试者工作特征曲线评估各模型诊断AD与aMCI的效能.结果 小波特征在可用于诊断AD与aMCI的最优海马纹理特征中占比最高.基于各算法的联合模型诊断测试集AD的曲线下面积(AUC)均最高(0.996、0.993、0.991),18F-FDG PET模型次之(0.941、0.941、0.967)而3D T1WI模型最低(0.801、0.801、0.750).基于LR、RF算法的联合模型诊断测试集aMCI的AUC 最高(0.967、0.992),18F-FDG PET 模型次之(0.951、0.971),3D T1WI 模型最低(0.833、0.824).基于 SVM 算法的联合模型与18F-FDG PET模型诊断测试集aMCI的AUC相同(0.951)并均高于3D T1WI模型(0.833).结论 PET/MRI海马纹理分析有助于诊断AD及aMCI;多模态联合诊断优于单模态,且具有良好稳定性.
Abstract
Objective To investigate the diagnostic value of PET/MRI hippocampal texture analysis for diagnosing Alzheimer disease(AD)and amnestic mild cognitive impairment(aMCI).Methods Data of 55 patients with AD(AD group),60 patients with aMCI(aMCI group)and 55 healthy controls(HC group)were retrospectively analyzed.The subjects were randomly divided into training and testing sets in a ratio of 7:3.Simultaneous PET/MRI was performed to obtain 3D T1WI and 18F-FDG PET.Texture features of ROI on both side hippocampus in training set were extracted.Logistic regression(LR),support vector machine(SVM)and random forest(RF)were used to construct 3D T1WI models,18F-FDG PET models and combined models,respectively.Receiver operating characteristic curves were drawn to evaluate the efficacy of the above models for diagnosing AD and aMCI.Results The wavelet features accounted for the most of the optimal hippocampal texture features for diagnosing AD and aMCI.Based on LR,SVM and RF algorithms,the area under the curve(AUC)of the combined models(0.996,0.993,0.991)were all the highest for diagnosing AD in testing set,followed by 18F-FDG PET models(0.941,0.941,0.967),and of single-modal model were the lowest(0.801,0.801,0.750).Based on LR and RF algorithms,AUC of the combined models(0.967,0.992)were highest for diagnosing aMCI in testing set,followed by 18F-FDG PET models(0.951,0.971),and the 3D T1WI models had the lowest AUC(0.833,0.824).Based on SVM algorithm,AUC(0.951)of combined model and of 18F-FDG PET model for diagnosing aMCI in testing set were the same,both higher than that of 3D T1WI model(0.833).Conclusion PET/MRI hippocampal texture analysis was helpful for diagnosing AD and aMCI.Multimodal combined diagnosis was superior to single-modal,with good robustness across different machine learning models.
基金项目
国家自然科学基金(82102010)
北京市科技新星计划(2021B00001609)
北京市科技新星计划(20220484177)
北京市科技计划(Z201100005520018)
NETL
NSTL
万方数据