摘要
机器学习能促进静息态功能磁共振成像(rf MRI)在癫痫中应用,尽管Pearson相关性的传统功能连接(FC)模型作为成像算法有较多报道,但其分类鲁棒性却少有研究。提出特异于健康人的癫痫患者FC指数模型,与FC在有监督机器学习分类敏感性和稳定性上进行比较,以期为提取癫痫患者功能影像学标记提供新算法。搜集20名结构像标记为海马阳性的内侧颞叶癫痫患者(各10名纳入左侧、右侧2组)和142名来自连接组学且与患者相同年龄段健康人的rf MRI数据;以健康人群为参照,构建个体患者FC特异性指数模型,为每个脑区功能打分;通过ROC敏感性分析曲线和曲线下面积(AUC)提取指数模型,对发作侧敏感脑区获得功能影像标记;以其指数作为特征向量,分别输入至概率神经网络和支持向量机,对患者发作侧分类;10次随机交叉验证分析稳定性,再分别对敏感脑区之间和患者之间的特征向量做线性相关性分析,以探求影响稳定性的内在原因。最后,用FC代替指数模型做同上处理,并比较两种功能连接模型的分类稳定性。结果显示,以FC为特征向量的AUC为0.76,而特异性指数的特征向量AUC为0.84,指数模型的分类敏感性高于FC。另外,FC的分类精度在25%~100%之间强烈波动,方差高达25.99%,且特征向量平均相关系数为0.67,相关性较强;而指数模型则在75%~100%之间较小波动,方差低至7.10%,且特征向量平均相关系数为0.28,相关性较小。在机器学习癫痫定侧中,静息态功能连接特异性指数模型表现出较强的分类鲁棒性,远优于传统模型,特征向量相关性较大可能是影响后者稳定性的主要原因。
Currently,machine learning has promoted the application of resting-state functional magnetic resonance imaging(rfMRI)in epilepsy,where the functional connectivity model of Pearson correlation(FC)has been widely applied as a traditional imaging algorithm.However,the classification stability of functional connectivity model is rarely studied in the machine learning.To address this issue,a FC-based index model specific to the healthy people was proposed in this work,the classification stability was studied by a random cross validation in the supervised machine learning models,and compared to the results of FC,aiming to provide a new algorithm in extracting FC features input into machine learning.The rfMRI data of a total of twenty patients of medial temporal lobe epilepsy with a positive indicator of hippocampus on structure MRI(equally involved in a group of left side and a group of right side),and a total of 142 healthy people from a connectome including Southwest Adult Lifespan Dataset(SALD)in the same age group were collected.A rfMRI FC-based index model was built up,specific to the healthy people,referred as FC-based specificity index model.Thus,every FC of each brain area in an individual patient could be scored,and the brain areas sensitive to paroxysmal side could be extracted by the ROC curve.The sensitivity analysis curve was taken as the functional bio-markers,whose indexes were assigned as the feature vectors to input into the supervised machine learning models such as probabilistic neural network(PNN)and support vector machine(SVM)to classify paroxysmal side.Additionally,the classification stability was validated by a random cross validation(10 times),and the linear correlation of feature vectors between sensitive brain areas and between patients were estimated to evaluate their interdependence,aiming to find out the underlying cause to affect the classification stability.Finally,the same procedures as above were fulfilled by the FC model instead of FC-based specificity index model,and the classificatio
作者
杨泽坤
葛曼玲
付晓璇
陈盛华
张夫一
郭志彤
张志强
Yang Zekun;Ge Manling;Fu Xiaoxuan;Chen Shenghua;Zhang Fuyi;Guo Zhitong;Zhang Zhiqiang(State Key Laboratory of Reliability and Intelligence of Electrical Equipment,Hebei University of Technology,Tianjin 300130,China;Key Laboratory of Electromagnetic Field and Electrical Apparatus Reliability of Hebei Province,Hebei University of Technology,Tianjin 300130,China;Department of Medical Imaging,Jinling Hospital,Nanjing University School of Medicine/General Hospital of Eastern Theater,Nanjing 210002,China)
出处
《中国生物医学工程学报》
CAS
CSCD
北大核心
2021年第5期521-530,共10页
Chinese Journal of Biomedical Engineering
基金
河北省研究生创新项目(CXZZSS2021034)
国家自然科学基金(81871345)
河北省省级科技计划项目(E2019202019)。
