Brain Functional Activity Detection Based On Neural Networks

Human brain is a huge system with the most complex structure. As the human beings evoluting, human brain is continuously evoluting. Its function is enriched dramatically, and has become the most complicated staff. And, brain imaging is one of the youngest and the most rapidly developing scientic fields. It emerges at the intersection of several of well-established disciplines. In the present study, focusing on the application and development of the neural network on fMRI data processing, we home discusseed the localization of functional activation. Meawhile, an innovative research of simultaneous epilepsy EEG-fRMI also successfully developed and primary results of clinical epileptic localization showed the beautiful effect.Some aspects of this dissertataion have been put forward:The spatial independent component analysis (sICA), as a neural network methodology, presented that it separated the different response of the complex visual-movement task into the spatial separated pattern. Our results showed that sICA method could separate two independent component patterns corresponded to each task in multi-task experiment and that the brain functional visual response is faster than movement response.The main objective of this section was that employing the sICA analyzed the EEG-fMRI data in 11 patients with focal epilepsy. Futher, we compared this approach with conventional multiple, fixed HRFs method. Our results showed that if the relationship asscioated with the BOLD response and the phenomenon of epileptic spikes for the scalp EEG recording is still uncertain, it is the better choice that the sICA as accurately identifying the epileptic lesion.Self-organizing mapping (SOM), as a supervised neural network, grouped image voxels into clusters, which used the similarity of their time course based upon the assumption that the pattern of activation had a structure and can be divided into several types of similar activations. In this section, the conventional Euclidean distance metric was replaced by the correlation distance metric in SOM to improve clustering and merging of neural nodes. To improve the use of spatial and temporal information in fMRI data, a new spatial distance and temporal correlation was introduced. It is more reasonable that taking advantage of the intrinsic feature of temporal and spatial information.Hidden Markov model (HMM), a special neural network method was applied in the fMRI data processing. In this part, a two-state HMM based on Bayesian probabilistic methodology was introduced to locate brain functional activation regions and to reveal state transition of the time course in analyzing event-related fMRI data. Log likelihood maps for all voxels were firstly calculated by first-order continuous density HMM, where the probability densities function was associated with simple on-off states to reveal activating regions. A state transition map for each given voxel was then characterized by the fMRI time course intensity change that described the neural response to the special stimuli.