目的 探讨fMRI数据中白噪声及去噪预处理方法对于采用FastICA算法检测人脑激活区精度的影响.方法 采用模拟和真实fMRI数据进行测试.实验重复50次,以比较未去噪、高斯平滑和两种正交小波方法去噪预处理后FastICA算法分离激活信号的能力.对不同阈值条件下检测的激活体素进行统计分析.用ROC曲线评价检测质量.结果 模拟实验显示,当fMRI数据的SNR为15 db时,去噪预处理能显著提高FastICA检测准确率;SNR为20 db时,去噪对提高FastICA检测准确率影响较小.采用不同的小波基及不同参数去噪处理结果差异较大,存在经db4小波去噪处理后激活区检测结果的敏感度和特异度反而降低的现象,但随着激活区减小.这种现象消失.结论 对于FastICA而言,小波去噪与传统的高斯平滑方法相比未显示出更好的敏感度和特异度;在低SNR情况下,有必要进行去噪预处理;在高SNR情况下,不恰当的去噪方法反而可导致FastICA检测精度降低.%Objective To investigate the impact of white noise and denoising preprocessing in fMRI on the accuracy of the human brain activation detection with FastICA algorithm. Methods Both simulated and real fMRI data were used to test the performance of FastICA. The experiments were repeated 50 times to compare the activation detection with FastICA before and after denoising with Gaussian smoothing and two orthogonal wavelets, respectively. Then statistical analysis was performed on the activated voxels which were detected under the different thresholds. Finally, ROC curves were used to evaluate the quality of detection. Results Simulation results showed that when SNR of fMRI data was 15 db, denoising preprocessing could significantly improve the detection accuracy of FastICA, whereas when the SNR was 20 db, denoising had little effect on the detection accuracy. The results were different when different wavelet bases and parameters for denoising were selected. After denoising by db4 wavelet, sometimes the sensitivity and specificity of the activation detection got worse in high SNR, but this phenomenon disappeared with the decrease of activation. Conclusion Compared with the traditional Gaussian smoothing method, wavelet denoising does not show better sensitivity and specificity for FastICA. It is necessary to denoise in low SNR, but in high SNR, inappropriate denoising method can lead to lower detection accuracy of FastICA.
展开▼
