基于不同分割方法的18F-FDG PET/MR影像组学鉴别帕金森病和多系统萎缩

18F-FDG hybrid PET/MR radiomics based on different segmentation methods for distinguishing Parkinson′s disease from multiple system atrophy

目的探索不同分割方法构建的18F-FDG PET/MR影像组学模型对鉴别帕金森病(PD)和多系统萎缩(MSA)诊断效能的影响。方法回顾性收集2017年12月至2019年6月间于华中科技大学同济医学院附属协和医院行18F-FDG PET/MR检查的PD及MSA患者共90例[男37例、女53例,年龄(55.8±9.5)岁],其中PD患者60例,MSA患者30例,按7∶3的比例随机分为训练集和验证集。采用自动标签功能解剖学(AAL)脑区模板匹配图像的自动脑区分割方法以及ITK-SNAP软件手动逐层分割方法勾画双侧尾状核及壳核作为ROI,分别从18F-FDG PET和T1加权成像(WI)中各提取1172个影像组学特征。采用最小冗余最大相关性(mRMR)和最小绝对收缩和选择算子(LASSO)算法对训练集进行特征降维并建立影像组学模型,同时采用十折交叉验证以减少模型过拟合。采用ROC曲线评价不同分割方法建立的影像组学模型在训练集及验证集中的鉴别诊断效能,并采用Delong检验比较其差异。结果训练集63例(42例PD,21例MSA),验证集27例(18例PD,9例MSA)。采用自动分割和手动分割所建立的影像组学模型(18F-FDGRadscore和T1WIRadscore)在训练集和验证集中,其Radscore值在PD组和MSA组之间差异均有统计学意义(z值:-5.15~-2.83,均P<0.05)。基于自动分割的18F-FDGRadscore和T1WIRadscore在训练集、验证集的ROC AUC分别为0.848、0.840和0.892、0.877;基于手动分割的两者的AUC分别为0.900、0.883和0.895、0.870;在训练集或验证集中,基于自动和手动分割方法所建立的影像组学模型的诊断效能之间的差异均无统计学意义(z值:0.04~0.77,均P>0.05)。结论基于自动分割和手动分割方法的18F-FDG PET/MR影像组学在鉴别PD和MSA中均有较好的诊断效能,但自动分割省时省力且可重复性较高,其在PD和MSA鉴别诊断中具有更大的潜力和实用价值。

Objective To explore the impact of different segmentation methods on differential diagnostic efficiency of 18F-FDG PET/MR radiomics to distinguish Parkinson′s disease(PD)from multiple system atrophy(MSA).Methods From December 2017 to June 2019,90 patients(60 with PD and 30 with MSA;37 males,53 females;age(55.8±9.5)years)who underwent 18F-FDG PET/MR in Union Hospital,Tongji Medical College,Huazhong University of Science and Technology were retrospectively collected.Patients were randomized to training set and validation set in a ratio of 7∶3.The bilateral putamina and caudate nuclei,as the ROIs,were segmented by automatic segmentation of brain regions based on anatomical automatic labeling(AAL)template and manual segmentation using ITK-SNAP software.A total of 1172 radiomics features were extracted from T1 weighted imaging(WI)and 18F-FDG PET images.The minimal redundancy maximal relevance(mRMR)and least absolute shrinkage and selection operator(LASSO)algorithm were used for features selection and radiomics signatures(Radscore)construction,with 10-fold cross-validation for preventing overfitting.The diagnostic performance of the models was assessed by ROC curve analysis,and the differences between models were calculated by Delong test.Results There were 63 cases in training set(42 PD,21 MSA)and 27 cases in validation set(18 PD,9 MSA).The Radscore values were significantly different between the PD group and the MSA group in all training set and validation set of radiomics models(18F-FDG_Radscore and T1WI_Radscore)based on automatic or manual segmentation methods(z values:from-5.15 to-2.83,all P<0.05).ROC curve analysis showed that AUCs of 18F-FDG_Radscore and T1WI_Radscore based on automatic segmentation in training and validation sets were 0.848,0.840 and 0.892,0.877,while AUCs were 0.900,0.883 and 0.895,0.870 based on manual segmentation.There were no significant differences in training and validation sets between Radiomics models based on different segmentation methods(z values:0.04-0.77,all P>0.05).Conclusions The 18F-FDG PET/MR radiomics models based on different segmentation methods achieve promising diagnostic efficacy for distinguishing PD from MSA.The radiomics analysis based on automatic segmentation shows greater potential and practical value in the differential diagnosis of PD and MSA in view of the advantages including time-saving,labor-saving,and high repeatability.