The human brain is an elegant, complicated and efficient system, which controls human emotion, mind, and behavior. The development of imaging technology has provided noninvasive approaches for investigating the structure and function of the human brain with the help of brain image analysis techniques. Brain image segmentation and registration, as fundamental issues in brain image analysis, have been hot research topics in recent years. Build upon the success of existing brain image segmentation and registration methods, we have proposed novel and effective brain image segmentation and registration methods, including: 1. A novel method is proposed for automatic segmentation of brain tumor images by leveraging both statistical image information and individual image information, which effectively alleviates the influence of the inconsistent image intensity information between training images and the image to be segmented. In particular, the statistical discriminative information provided by supervised learning is first incorporated with the multi-scale structure information of the image to be segmented to identify the initial label information for tumor and normal region robustly. And then the graph based label information propagation technique is used to segment the tumor region accurately. The validation on both simulated and clinical brain tumor images has demonstrated the effectiveness and superiority of the proposed method. 2. A new strategy for multi-scale functional brain parcellation is proposed with the number of regions at different spatial scales estimated in a data-driven way, and the consistency of inter-subject brain functional network properties are used to identify the proper spatial scales for the parcellation. In particular, functional brain images are modeled as weighted graph, with graph nodes representing image voxels and graph edge measuring the inter-voxel functional similarity. By iteratively selecting functionally representative nodes and weighted aggregation of functionally similar nodes, the whole brain is parcellated at different scales hierarchically. Global functional similarity constraints are incorporated in the procedure of node aggregation to reduce the adverse effects of ignoring the size information of different functional brain regions in the spatially local aggregation procedure. Moreover, subject specific functional parcellation is achieved based on the group functional parcellation results and subject specific functi...
修改评论