基于磁敏感加权成像的黑质影像组学诊断帕金森病的研究

The study of diagnosing the Parkinson's disease based on substantia nigra radiomics on susceptibility-weighted imaging

目的 探讨磁敏感加权成像(susceptibility-weighted imaging, SWI)的黑质影像组学及机器学习方法在帕金森病(Parkinson’s disease, PD)诊断中的价值。材料与方法 回顾性分析80例早期PD患者和80例健康受试者的SWI图像。采用ITK-SNAP软件对SWI上黑质区域进行感兴趣区勾画,并提取和筛选影像组学特征。应用5种机器学习方法(支持向量机、逻辑回归分析、随机森林、贝叶斯、K近邻)构建PD诊断模型,选择诊断效能最好且最稳定的模型进行验证,并与人工识别燕尾征的诊断效能进行比较。结果 共筛选出7个与PD密切相关的影像组学特征。在训练集中,逻辑回归分析模型的诊断效能最好(AUC=0.975)且最稳定(AUC的相对标准差为4%)。在测试集中,逻辑回归分析模型诊断PD的AUC为0.938,敏感度为83.3%、特异度为95.8%,其诊断效能明显优于人工识别(Z=2.241, P=0.025)。结论 基于SWI黑质的影像组学特征构建的逻辑回归分析模型可准确诊断PD,为临床早期干预治疗提供影像指导。

Objective:To explore the value of the substantia nigra radiomics on susceptibility-weighted imaging(SWI)and machine learning in diagnosing the Parkinson's disease(PD).Materials and Methods:SWI images of 80 early PD patients and 80 healthy controls were collected.The ITK-SNAP software was used to delineate the regions of interest in substantia nigra on SWI,and extract and screen radiomics.Five machine learning methods(support vector machine,logistic regression analysis,random forest,Bayesian,K-nearest neighbor)were used to build the diagnosis model for Parkinson's disease,and the model with best diagnosis efficiency and most stable was selected for verification,and compared with the diagnosis efficiency of visual analysis of swallow tail sign.Results:A total of 7 radiomic features closely related to Parkinson's disease were screened.In the training set,the diagnostic performance of the logistic regression analysis model is the best(AUC=0.975)and the most stable(relative standard deviations of AUC:4%).In the test set,the AUC of logistic regression analysis model for diagnosing PD was 0.938,the sensitivity was 83.3%,and the specificity was 95.8%,which was significantly better than that of visual analysis(Z=2.241,P=0.025).Conclusions:A logistic regression analysis model based on substantia nigra radiomics on SWI can accurately diagnose PD and provide image guidance for early intervention treatment in clinical.