The Application Research On Clustering Methods For FMRI Data Analysis

Functional magnetic resonance imaging (fMRI) is a neuroimaging technique which is used to study the human brain's function in response to stimulus, and is the fastest growing and most effective as a technology measurement of brain activity now. The functional magnetic resonance imaging combined of three factors-function, image and anatomical, and detect active regions for feeling, motion and cognition by analysis of blood oxygenation level-dependent signal changes. The function magnetic resonance image has many advantages, such as non-invasiveness, radiation-free, high temporal and spatial resolution, the operation can be repeated and so on. Therefore, functional magnetic resonance imaging as the preferred method of functional brain imaging has been more extensive applications.With functional magnetic resonance imaging technology, a number of methods for brain fMRI data processing came into being and made a lot of results. These include general linear model (GLM), ICA, clustering method.In this paper we used a non-supervised clustering method applied to fMRI data analysis. The main aspect of a clustering method for a particular problem is to choose a criterion to determine the members of the same category. For fMRI data, whether a special voxel belongs to one active voxels group, is amplitude-independent, which is only allied to the time waveform. That is to say the shape of the vector clouds would be cylindrical instead of the more common spherical clouds around class extracted by other clustering method. In this paper we have used a cylindrical standard for clustering, and built the probability distribution for the model, and using EM algorithm got distinguishable activation time signature vectors based on maximum likelihood estimation rule. Maximum likelihood criteria could not determine the optimal number of categories, as the maximum number is non-decreasing function of the likelihood function. Our algorithm also needed to estimate the optimal number of categories, so we added penalty factor to likelihood function based on Minimum Descriptions Length criteria, and computed the value of it to get the optimal number of categories.After getting signature time vectors, the activation maps in response to each activity should be generated. FMRI activation maps could indicate relationship between the different desired activities. The proposed method generated activation maps was Optimal Linear Transformation (OLT), which was an image analysis technique of feature space. OLT could maximize the SNR of regions with the desired activity and remove regions with the undesired activities or inactive regions.In this paper, we used the ideal hemodynamic response to produce synthetic fMRI data. We analyzed the synthetic data using the proposed clustering method and GLM based analysis in AFNI separately. Then we compared their performance using receiver operating characteristic curves (ROC), which is a approach widely used in evaluation of medical diagnosis test in statistical aspects. The results indicate that, the proposed method was more accurate for brain activation detection over the GLM method in our simulation. The proposed method is applied to the data analysis of a blocked fMRI experiment, and we compared the results with ICA.