Research On Group Analysis Technology Of Functional Magnetic Resonance Imaging Data

With the rapid development of brain imaging technology,increasing amount of data are being available.The group analysis of multiple datasets is becoming an important subject in brain imaging data analysis.Group analysis focuses on exploring similarity among multiple datasets and diversity of each dataset.Based on big data,group analysis can obtain reliable results.At present,there exist two issues in group analysis technology,including the challenge of order determination for multiple dimensionality reductions and the lack of prior information incorporated into analysis.In this study,group analysis methods are improved to solve these problems.The main work of this paper is summarized as follows:The dimensional optimization concept was proposed and applied to improve the group analysis performance.The dimensionality reduced dimensions are rotated to the optimized direction to improve the accuracy and robustness of the group analysis method.Due to the high dimensionality of multiple brain imaging datasets,dimensionality reduction is generally performed for multiple times before blind source separation.And the number of retained dimensionality dramatically influences the validity and robustness of blind source separation models.Inspired by the idea of signalintensity-maximizing technology,the retained dimensions are rotated to the optimized direction so that the rotated components have the most significant intensity and smoothness.Thus the proposed method can make full use of useful information.Results on simulated data showed that involving subject-level dimensional optimization can significantly enhance the accuracy of group analysis,and involving group-level dimensional optimization can significantly improve the robustness of the method.Experimental results on real data indicated that,the new method extracted more reliable task-related components and showed superior performance in practical applications.The group analysis model is improved by involving prior information.The temporal feature of the experimental paradigm is used to impose constraint to the model,and the activation detection power of the group analysis method is improved.Because the magnitude of the task/stimulus-related signal changes only after the occurrence of task/stimulus,there exists significant BOLD-level difference in the activated regions.Based on this feature,weak constraints are applied to the canonical correlation analysis model to improve the activation detection power.Results on simulated data showed that,compared with the traditional algorithms without prior information constraints or with strong constraints,the proposed algorithm achieved higher accuracy and stability in noise conditions.And the new method showed higher detection power in separating each single task-related activation from multi-task brain imaging data.During the experiments on real task-state brain imaging data,the activations obtained by the proposed method were in line with expectation,while the taskrelated time courses extracted by the new method achieved higher correspondence with the paradigm and better consistency among subjects.And the new method successfully detected transiently task-related components from real data.