Application Of Mean-shift Clustering Based On Multi-feature Space To Functional MRI

Functional magnetic resonance imaging(fMRI) is a powerful tool for brain science. Due to the existence of noises, there is little difference between signal and noise in the region of interest, and it is difficult to extract fMRI signal directly. In this paper, we proposed two novel f MRI activation detection algorithms to overcome the disadvantages of current detection methods by utilizing the spatial features and temporal features.At first, on the basis of temporal features of fMRI data, we proposed a mean-shift clustering method based on frequency domain(FD-MSC) to detect the activated region in fMRI data. The proposed method combined temporal features with mean-shift clustering. The frequency domain was obtained by fast Fourier transform to construct a feature space, and then mean-shift clustering was adopted to detect the active region of fMRI based on the feature space.Second, on the basis of spatial features of fMRI data, we proposed a mean-shift clustering method based on voxel’s neighborhood(VN-MSC) for fMRI activated detection. The proposed method combined spatial features with mean-shift clustering. The voxel’s neighborhood obtained by cross-correlation analysis method was used to construct two dimensional feature spaces, which efficiently integrates fMRI data voxel’s neighborhood information. Finally, mean-shift clustering based on this feature space was adopted to detect active region of fMRI.The two proposed methods were applied to simulated data and real fMRI data for evaluation. The results of the evaluation using simulated data showed that the sensitivity and specificity of FD-MSC and VN-MSC with appropriate kernel size were better than cross-correlation analysis(CCA) and CCA plus cluster analysis(CCA+CA). The evaluation results of real fMRI data showed that the FD-MSC and VN-MSC method were consistent to other two methods(CCA, CCA+CA) in accuracy, while the detection region of the proposed methods was more complete. According to the evaluation results, the FD-MSC method makes full use of the temporal features of fMRI data and it is suitable for detection of the signal with stronger frequency domain. Meanwhile, the VN-MSC method makes full use of the spatial features of data and it is suitable for the detection of signal with stronger spatial correlation. The two proposed methods both have good robustness and high sensitivity, and they are suitable for fMRI activation analysis.