摘要
目的:为了明确白质高信号的危险因素和发病机制,基于迁移学习和空间协方差模型研究脑萎缩和白质高信号的相互关系及高血压对这种关系的影响。方法:选取美国亚利桑那大学附属医院41名健康中老年人(健康对照组)和41名高血压中老年患者(高血压组)MRI数据。首先,采用基于种子点的模糊聚类算法测量白质高信号的体积。其次,基于上百万张图片预训练得到的卷积神经网络AlexNet采用迁移学习的方法从灰质密度图中提取大脑形态特征。最后,采用基于主成分分析方法的尺度子配置模型来获取与脑白质病变相关的形态学空间协方差模式。结果:与灰度密度图相比,深度学习模型对影像特征具有更好的表达,通过迁移学习提取的特征具有更高的模型拟合优度。与健康对照组相比,高血压组患者的脑部影像对这种特征模式具有更强的表达。结论:基于迁移学习和空间协方差方法可以较好地应用于脑萎缩和白质高信号间关系的研究中。
Objective To examine the relationship between brain atrophy and cerebral small vessel disease and the influence of hypertension on this relationship with transfer learning and spatial covariance method.Methods Totally 41 healty millde-aged and elderly people(enrolled into a control group)and 41 middle-aged and elderly hypertension patients(hypertension group)from the affiliated hospital of The University of Arizona had their MRI data involved into the study.First,a fuzzy clustering method based on seed points was used to measure the volume of white matter hyperintensity(WMH).Second,a convolutional neural network,AlexNet,pretrained on around one million images was used to extract morphometric features from gray matter map.Finally,Scaled Subprofile Model(SSM),a principal components analysis(PCA)-based method was applied to characteristic spatial covariance pattern associated with white matter lesion.Results Compared with gray matter map,deep learning model had a better representation of image features;the features extracted from transfer learning had a higher goodness of fit.It was also found that the pattern was expressed more in the hypertension group than the normotensives.Conclusion The proposed method based on transfer learning and spatial covariance modeling can be used to study the relationship between brain atrophy and white matter hyperintensity.
作者
吴玉超
林岚
宋爽
吴水才
WU Yu-chao;LIN Lan;SONG Shuang;WU Shui-cai(College of Life Science and Bioengineering,Beijing University of Technology,Beijing 100124,China)
出处
《医疗卫生装备》
CAS
2019年第1期10-13,25,共5页
Chinese Medical Equipment Journal
基金
国家科技支撑计划课题(2015BAI02B03)
北京市自然科学基金-海淀原始创新联合基金资助项目(L182010)
北京市教委科技计划项目(KM201810005033)
