摘要
为解决受限玻尔兹曼机(RBM)在功能磁共振成像(fMRI)脑功能连通性检测中遇到的体素数量过多和模型参数难以选择的问题,提出一种结合主成分分析(PCA)和Bootstrap区间估计的受限玻尔兹曼机方法,选出fMRI数据中的部分体素,从而削减体素数量。以经体素削减处理后剩余体素的时间过程作为样本,采用改进的学习算法训练RBM,根据模型权重参数重建脑功能网络空间图谱。实验结果表明,在单被试fMRI脑功能联通性检测中,基于RBM的方法在空间域和时间域中的分析结果明显优于稀疏近似联合受限玻尔兹曼机方法。基于RBM的方法和Infomax ICA方法的空间域ROC曲线非常接近,但前者在时间域上的时间过程与实验刺激BLOCK的相关性更高。实验结果表明,基于RBM的方法能够有效地降低样本中的体素数量和模型参数选择的复杂度,提高RBM在fMRI数据分析中的性能。
In order to solve problems of the Restricted Boltzmann Machine（RBM） encountered in functional Magnetic Resonance Imaging （fMRI） brain functional connectivity detection,including excessive number of voxels and difficult to choose model parameters,this paper proposes a RBM method,which combines Principal Component Analysis （PCA） with Bootstrap interval estimation.It is introduced to select a part of voxels from fMRI data,so as to reduce the number of voxels,taking the time courses of remained voxels as sample and training RBM with improved learning algorithm.According to the weight parameters of the model,the spatial maps of the brain functional networks are constructed.Experimental results show that,in brain functional connectivity detection of single subject,the analysis results of RBM method in spatial domain and time domain are significantly better than sparse approximation joint restricted Boltzmann machine method.The spatial domain ROC curve of the RBM method and Infomax ICA are very close,but the temporal time course of the former is more relevant to the BLOCK of experimental stimulation.Experimental results show that the RBM method can effectively reduce the number of the voxels in sample and complexity of the model parameters,thereby improves the performance of RBM in fMRI data analysis.
出处
《计算机工程》
CAS
CSCD
北大核心
2017年第1期231-236,共6页
Computer Engineering
基金
国家自然科学基金(31470954)
上海市科委基金(14590501700)
