【作者】 李辉；

【导师】 朱文浩；

【作者基本信息】 华中科技大学 ， 神经病学（专业学位）， 2023， 硕士

【摘要】 目的:脑小血管病(small vessel disease,SVD)和多发性硬化(multiple sclerosis,MS)均以脑白质高信号(white matter hyperintensity,WMH)为主要影像学表现,且临床表现常缺乏特异性,如何对二者进行快速、准确的诊断是目前亟待解决的临床问题。本研究拟利用MRI平扫上的定性影像学资料和基本临床信息,基于机器学习和列线图模型实现对SVD和MS的快速精准鉴别。方法:本研究纳入161例SVD和121例MS住院患者。收集患者基本临床资料,并对MRI平扫中的一系列相关影像学特征进行定性评估。对特征进行降维处理后,利用机器学习方法(Logistic回归、支持向量机、随机森林、XGBOOST),分别基于纯影像学特征,和临床+影像学特征探索SVD和MS的有效分类模型,并在所有被试及40-65岁年龄段的被试(MS和SVD患者的年龄分布多重合于该年龄段)中分别检验分类器的效能。进一步地,基于Logistic回归分类器中各特征的权重,绘制列线图及构建SVD和MS鉴别诊断的临床量表。结果:基于纯影像特征,共筛选出10个特征用于构建最终的MS和SVD分类模型,分别为:幕上T1W1黑洞征、边界清楚的圆形/卵圆形病灶、近皮层WMH、脑桥周围部WMH和下颞叶WMH(倾向于诊断MS);腔隙(lacune,LA)、LA>3个、基底节区扩大的血管周围间隙(enlarged perivascular space,PVS)>1级、基底节区WMH、外囊区WMH(倾向于诊断SVD)。四种分类器(Logistic回归、支持向量机、随机森林、XGBOOST)在所有被试和40岁-65岁年龄段间的被试中均能实现对MS和SVD的准确区分。基于临床特征+影像特征模型,共筛选出9个特征用于构建最终的MS和SVD分类模型,分别为:年龄、高血压、幕上T1W1黑洞征、边界清楚的圆形/卵圆形病灶、下颞叶WMH、延髓病灶、LA、LA>3个、外囊区WMH。上述四种分类器在所有被试和40岁-65岁年龄段间的被试中均能实现对MS和SVD的准确区分。在以纯影像特征构建的量表中(0-60分),总分>27分倾向于诊断为SVD,否则诊断为MS,该量表在所有被试和40岁-65岁年龄段被试中的分类AUC分别为0.884/0.877(训练集/测试集),0.844/0.818(训练集/测试集)。在以临床特征+影像特征构建的量表中(0-200分),总分>145分,诊断为SVD,否则诊断为MS;该量表在所有被试和40岁-65岁年龄段被试中的分类AUC分别为0.951/0.965(训练集/测试集),0.915/0.939(训练集/测试集)。结论:基于MRI平扫上的纯影像学特征能实现成人MS和SVD的良好区分,在所有被试和40-65岁的被试中均具有良好而稳定的分类效能;利用临床+影像特征能进一步地提升整体成年人群中的诊断准确度。基于筛选特征构建的临床量表可实现临床上对SVD和MS的快速准确鉴别。更多还原

【Abstract】 Objective: Both small vessel disease(SVD)and multiple sclerosis(MS)are mainly characterized by high white matter hyperintensity,whereas their clinical manifestations often lack of specificity.How to differentiate them quickly and accurately is a clinical problem that needs to be solved urgently.In this study,qualitative imaging data from brain magnetic resonance imaging(MRI)and basic clinical information were used to achieve rapid and accurate identification of SVD and MS based on machine learning and Nomogram.Methods: In this study,161 patients with SVD and 121 patients with MS were enrolled.The basic clinical data of patients were collected,and the relevant imaging features in MRI plain scan were qualitatively evaluated.After feature dimension reduction,machine learning methods(logistic regression,support vector machine,random forest and XGBOOST)were used to explore effective classification models for SVD and MS based on pure imaging features and clinical + imaging features,respectively.The efficiency of the classifiers in all subjects and in the 40-65 age group(the age distribution of MS and SVD patients mainly overlaps in the 40-65 age group)was both examined.Further,based on the weight of each feature in the Logistic regression classifier,Nomograms were drawn and clinical scales for differential diagnosis of SVD and MS were constructed.Results: Based on the pure imaging features,a total of 10 features were selected to construct the MS and SVD classification models,including supratentorial T1W1 black holes,welldefined circular/oval lesions,juxtacortical lesions,peripons lesions,and inferior temporal lobe involvement(which tended to diagnose MS),lacunes,lacunes>3,basal ganglia PVS>1,basal ganglia lesions and external capsular WMH(which tended to diagnose SVD).Four classifiers(Logistic regression,support vector machine,random forest and XGBOOST)were able to accurately distinguish MS from SVD in all subjects and those aged 40-65 years.Based on the clinical features + imaging features model,a total of 9 features were selected to construct the final MS and SVD classification models,including age,hypertension,supratentorial T1W1 black holes,well-defined circular/oval lesions,inferior temporal lobe lesions,medulla lesions,lacunes,lacunes>3,and external capsular WMH.Four classifiers(Logistic regression,support vector machine,random forest and XGBOOST)were able to accurately distinguish MS from SVD in all subjects and those aged 40-65 years.In the scale constructed with pure imaging features(0-60 points),that the total score>27 points tends to diagnose SVD,while that the total score≤27 points tends to diagnose MS.The classification AUC of this scale for all subjects and those aged 40-65 years old were 0.884/0.877(training set/ testing set),0.844/0.818(training set/testing set),respectively;In the scale constructed by clinical features + image features(0-200 points),that the total score>145 points tends to diagnose SVD,otherwise it tends to diagnose MS.The classification AUC of the scales in all subjects and those aged 40-65 years old were 0.951/0.965(training set/testing set)and0.915/0.939(training set/ testing set),respectively.Conclusions: The classifiers based on qualitative imaging features can well distinguish MS and SVD patients,which achieved a good and stable classification efficiency in both all subjects and 40-65 years subjects.The classifiers based on clinical features + imaging features can further improve diagnostic accuracy in the overall adult population.The clinical scale constructed based on screening features can realize the rapid and accurate identification of SVD and MS in clinic.更多还原