基于体素邻域信息的均值漂移聚类算法检测fMRI激活区

Detecting fMRI activation by mean-shift clustering method based on voxel neighborhood information

为了提高fMRI激活区提取算法的抗噪能力及精确性,提出了一种基于体素邻域信息的均值漂移聚类算法.采用互相关分析方法计算每个体素的时间序列与刺激函数的相关系数,并计算该体素的时间序列与邻域中体素的时间序列的相关系数,以这2种相关系数构建有效整合体素邻域信息的二维特征空间.再用均值漂移算法对此特征空间进行聚类搜索,完成对脑神经活动区域的检测.利用仿真数据和实际fMRI数据对算法进行测试.仿真数据测试结果表明,当选定合适的核宽,无论激活区域大小,所提出算法的敏感性和特异性均优于较传统的互相关分析算法和互相关聚类算法.实际fMRI数据测试结果显示,所提出算法与其他2种算法的结果具有良好的一致性,而所提出算法的检测区域更完整.

To improve the anti-noise capacity and precision of fMRI activation area detecting method, a mean-shift clustering method was proposed based on voxel neighborhood （VN-MSC）. A two-dimensional feature space was constructed for VN-MSC based on temporal properties of a yoKel and its neighbors. The correlation coefficient between MRI time-series and stimulation response function of each yoKel was calculated by cross-correlation analysis. The correlation coefficient between time-series of a voxel and the neighboring yoKels was also calculated by the same method. Based on the feature space, a mean-shift clustering was adopted to detect active region of fMRI to obtain simulated and real fMRI data. The VN- MSC method was tested by simulation data and practical fMRI data. The results show that the sensitivity and specificity of MSC-VN technique are better than those of the traditional cross-correlation analysis （CCA） and CCA plus cluster analysis in any activation area size when the kernel size is appropriate. The results of real tMRI data demonstrate that MSC-VN and other two methods have good consistency in accuracy, while the detection region of the proposed method is more complete.