Brain parcellation based on resting state fMRI data aims to automatically parcellate the brain regions into subunits with distinct functions using clustering algorithms. The brain parcellation is typically achieved by clustering brain image voxels according to a similarity measure defined by functional connectivity between functional signals. Most of the existing methods for brain parcellation adopt unsupervised clustering algorithms, e.g., K-means, hierarchical clustering, and normalized cut, which are not able to take into consideration any anantomical knowledge available. The present study has proposed a semi-supervised clustering method with a prior information to identify reliable boundaries between functional subunits for achieving a robust brain parcellation. The thesis mainly consists of following 2 parts. (1)A semi-supervised clustering method is proposed to achieve the brain parcellation guided by a prior information. The proposed method takes findings of the existing cytoarchitectonic studies, structural connectivity studies, and functional connectivity studies as a prior information to guide the brain parcellation for obtaining more robust parcellation results. Experiment results based on resting state fMRI data of 106 normal subjects have demonstrated that our method is able to successfully divide the cingulate cortex into six distinct functional subunits at an individual subject level. The robustness of our method has been validated using random initialization, and the validity and relibility of the pacellation results have been verified using functional connectivity pattern analysis and meta analysis. (2)An improved semi-supervised clustering algorithm is proposed to adopte spatial neighborhood information in the parcellation for obtaining smooth parcellation results. The proposed method iteratively propagates the label information of initial labeled voxels to their spatially nearest neighboring voxels with high similarity to them, facilitating smooth and reliable parcellation. Experiment resutls based on resting state fMRI data of 20 normal subjects have demonstrated that our method is able to divide the Broca area into two functionally distinct subunits. The comparison with established cytoarchitectonic data and parcellation results obtained by an unsupervised clustering algorithm has demonstrated our method is able to get more more accrute results.
修改评论