基于磁敏感加权成像的深度学习算法模型分析脑浅静脉特征与皮层下深部白质病变相关性

Susceptibility-weighted imaging joint deep learning to explore the potential association between superficial cerebral veins and deep white matter lesions

目的 基于磁敏感加权成像(susceptibility-weighted imaging,SWI)图像建立深度学习算法模型分析脑浅静脉(superficial cerebral veins,SCV)形态学参数与皮层下深部白质病变(deep white matter lesions,DWML)之间的关联。材料与方法 招募健康志愿者364名,所有受试者行头颅常规序列及SWI扫描。随机选取其中200名受试者的SWI图像重建得到最小密度投影(minimum intensity projection,MinIP)图像,采用深度学习算法分析MinIP图像,建立SCV形态学特征的自动量化模型。采用该模型分析剩余164名受试者双侧大脑半球SCV的血管直径、曲率等形态学参数。根据有无皮层下DWML将164名受试者分为两组,分别为有皮层下DWML组(共53人)和无皮层下DWML (non-deep white matter lesions,N-DWML)组(共111人)。采用Mann-Whitney U检验对DWML组与N-DWML组的SCV形态学参数进行分析。对差异有统计学意义的指标和皮层下深部白质病变建立二元Logistic回归模型。结果 用于分析SCV形态学特征的深度学习算法模型在训练集的平均准确率为98.19%,在验证集的平均准确率为98.02%,测试集的平均准确率为98.03%。双侧大脑半球SCV的各项量化指标在年龄及受教育年限的分布中差异均无统计学意义(P>0.05),但在性别中分布中差异有统计学意义,表现为男性双侧大脑半球SCV血管数量均显著多于女性(右侧P=0.004,左侧P<0.001)。DWML组和N-DWML组在双侧大脑半球的SCV直径及数量中差异均无统计学意义(P>0.05),DWML组的SCV曲率较N-DWML组显著增大(P<0.001)。右侧大脑半球SCV曲率与DWML的发生呈显著正相关(回归系数为2.035,P=0.015)。结论 SCV形态学改变与DWML之间存在潜在关联,SCV曲率改变可能通过影响局部血流动力学改变从而导致DWML的发生。

Objective:To explore the relationship between superficial cerebral veins’(SCV) morphological parameters and deep white matter lesions(DWML) through establishing a deep learning algorithm model based on susceptibility-weighted imaging(SWI).Materials and Methods:Three hundred and sixty-four healthy volunteers were recruited.All subjects underwent a routine head scan and SWI,SWI images of 200 subjects were randomly selected to reconstruct the minimum intensity projection(MinIP) image.The deep learning algorithms were used to analyze the MinIP images to establish the automatic quantification model of SCV morphological features.One hundred and sixty-four subjects were used for analyzing the morphological features of SCV in the bilateral cerebral hemispheres including diameter,tortuosity index(TI),and so on by using the deep learning models.According to whether there are DWML,the 164 subjects were divided into two groups including the group of DWML comprised 53 subjects and the group of no-deep white matter lesions(N-DWML) comprised 111subjects.The Mann-Whitney U test was used to compare the quantitative indicators of SCV between the group of DWML and N-DWML,and a binary logistic regression model was established for significant difference indicators and DWML.Results:The training set’s average accuracy rate of the deep learning algorithms model for analysis of SCV morphological features was 98.19%,the validation set’s average accuracy rate was 98.02%,and the test set’s average accuracy rate was98.03%.There were no differences in the distribution of age and years of education in the quantitative indicators of SCV in bilateral cerebral hemispheres(P>0.05),but there were significant differences in the distribution of gender,which showed that the number of SCV in males was significantly more than that of females(Right:P=0.004,Left:P<0.001).No significant difference was found in the diameter and number of SCV in bilateral cerebral hemispheres between DWML group and N-DWML group(P>0.05).However,the TI of SCV in DWML group was significantly larger than N-DWML group(P<0.001).The TI of SCV in right cerebral hemisphere was significantly correlated with the occurrence of DWML(regression coefficient=2.035,P=0.015).Conclusions:There is a potential correlation between the morphological changes of the SCV and the DWML.The changes in TI of the SCV may affect the local hemodynamic changes and lead to the occurrence of DWML.