颅内磁共振血管壁成像质量对影像组学模型鉴别症状性斑块的影响

The influence of the image quality of intracranial magnetic resonance vessel wall imaging on the identification of symptomatic plaques by radiomics model

目的:探讨颅内动脉MR血管壁成像(MR-VWI)的图像质量对基于影像组学特征构建症状性斑块预测模型的影响。方法:回顾性搜集因颅内动脉粥样硬化行MR-VWI检查且因图像质量不佳而即刻重复扫描的病例作为本研究的第一部分。MR-VWI序列包括全脑3D SPACE T_(1)WI平扫和增强,目标血管(单侧大脑中动脉或基底动脉)的2D TSE T_(2)WI。拟分析100个常用的影像组学特征,包括形状特征9个,一阶梯度特征18个,纹理特征73个。通过比较重复扫描的斑块影像组学特征,筛选出易受图像质量影响的不稳定特征。回顾性搜集因颅内动脉粥样硬化狭窄行MR-VWI检查的病例作为本研究的第二部分。首先,基于全部病例的斑块影像组学特征构建症状性斑块的预测模型(模型A);其次,剔除所有不稳定特征后构建预测模型(模型B);最后,剔除图像质量不佳者,构建预测模型(模型C)。重复扫描影像组学特征的比较,采用Wilcoxon符号秩和检验。预测模型的诊断效能采用受试者工作特征曲线(ROC曲线)进行分析,诊断效能高低主要通过曲线下面积(AUC)体现,以DeLong检验比较不同模型效能的差异。结果:第一部分纳入24例患者,发现3D SPACE T_(1)WI序列的形状特征和一阶梯度特征均无不稳定特征,纹理特征中的不稳定特征仅占1/73。3D SPACE T_(1)WI增强序列的形状特征、一阶梯度特征以及纹理特征中的不稳定特征占比分别是3/9、2/18和17/73。2D TSE T_(2)WI序列的形状特征、一阶梯度特征以及纹理特征中的不稳定特征占比分别是2/9、2/18和19/73。第二部分,102例患者的120个斑块纳入分析,其中症状性斑块51个,无症状斑块69个。模型A的AUC为0.708±0.022;与第一部的不稳定特征比对,模型A筛选出的8个影像组学特征中有3个不稳定特征。模型B的AUC为0.740±0.007。模型C预测颅内症状性斑块的预测效能最好,AUC为0.758±0.013;与第一部分的不稳定特征比对,模型C的8个影像组学特征均为稳定特征。结论:基于MR-VWI影像组学构建颅内症状性斑块的预测模型时,图像质量的优劣对斑块影像组学特征的有一定影响,图像质量越好则预测效能越高。

Objective:To investigate the effect of MR vessel wall imaging(MR-VWI)image quality on the construction of a symptomatic plaque prediction model based on radiomics signature.Methods:In the first part,patients who underwent MR-VWI examination for intracranial atherosclerosis were immediately re-scanned due to poor image quality and were retrospectively collected.MR-VWI sequences include whole-brain 3D SPACE T_(1)WI plain and enhanced scans,and 2D TSE T_(2)WI of the target vessel(unilateral middle cerebral artery or basilar artery).One hundred common radiomics features were analyzed,including 9 shape features,18 first-order gradient features,and 73 texture features.By comparing the plaque radiomics features of repeated scanning,the unstable features that were easily affected by image quality were screened out.In the second part,the cases of intracranial atherosclerotic stenosis who underwent MR-VWI examination were retrospectively collected.Firstly,a predictive model of symptomatic plaques was constructed based on the radiomics features of all cases in PartⅡ(Model A).Secondly,a prediction model(Model B)is constructed after removing all the unstable features.Finally,the cases with poor image quality were eliminated and a prediction model(Model C)was constructed.The Wilcoxon signed rank sum test was used to compare the radiomics features of repeated scan images.The diagnostic efficiency of the predictive model of symptomatic plaques based on radiomics features was analyzed using the receiver operating characteristic curve(ROC curve),and the diagnostic efficiency was mainly reflected by the area under ROC curve(AUC).The DeLong test was used to compare the efficacy of different models.Results:In the first part,a total of 24 cases were collected for immediate repeat scanning due to poor image quality.The results showed that shape features and first-order gradient features of the 3D SPACE T_(1)WI sequence were all stable,and the unstable features of texture features only account for 1/73.The unstable features of shape features,first-order gradient features,and texture features of 3D SPACE T_(1)WI enhanced sequences accounted for 3/9,2/18,and 17/73,respectively.The unstable features of shape features,first-order gradient features,and texture features of 2D TSE T_(2)WI sequences accounted for 2/9,2/18,and 19/73,respectively.In the second part,120 plaques from 102 patients were included in the analysis,including 51 symptomatic plaques and 69 asymptomatic plaques.The AUC of model A is 0.708±0.022.Compared with the unstable features in the first part,there are 3 unstable features in the 8 radiomics features selected by model A.The AUC of Model B is 0.740±0.007.Model C showed the best performance in predicting intracranial culprit plaques(AUC=0.758±0.013).Compared with the unstable features in the first part,all 8 radiomics features of model C are stable.Conclusion:When the prediction model of intracranial symptomatic plaques is constructed based on MR-VWI radiomics,the image quality has a certain influence on the plaque radiomics signature,and the better the image quality,the higher the prediction efficiency.