Research On Brain Computer Interface Based On Real-time Functional Magnetic Resonance Imaging

Brain computer interface (BCI) interprets the brain activities into computer commands to control external devices by decoding the brain activities in the brain signals, thus builds a connection path between the brain and external devices, which allows to operate external devices without using language or body actions. Brain-machine interface has become a hot research spot involving in multiple areas, such as biology, medicine, computer science, communication and so on. As a best noninvasive and repeatable means of measuring advanced brain functional activity, functional magnetic resonance imaging (fMRI) is currently one of the major methods to decode the human brain activities. The development of real-time functional Magnetic Resonance Imaging (rt-fMRI) has broken through the traditional offline fMRI data processing and analysis methods, makes it possible to observe brain activity in real time. Therefore researches on how to build a stable and reliable rt-fMRI system and rt-fMRI based advanced BCI systems are of significant theoretical and practical values.This paper explores some key issues in order to meet the requirements of the processing speed, reliability and accuracy based on the rtfMRI-BCI technology, such as the design method of the rt-fMRI based BCI system, the robust online activation detection method, the training and online update method of the BCI classifer. The main work is as follows:1. Based on the requirements of the rtfMRI-BCI, a modular design method for rt-fMRI based BCI system are proposed. A rtfMRI based BCI platform is built, through optimizing the performance of the core sub-module and functional extensions. For the real-time data access issue on different MRI equipment, a data acquisition method based on file system monitoring is proposed, which can improve the generality and practicality of the data acquisition module. Based on the functional requirements of online brain state classification and multiple forms of feedback, real-time data processing module and feedback module are designed and implemented, which are capable to support functional extensions.The test result for the built rtfMRI-BCI system shows that the system can complete the real-time data acquisition, head movement correction, online activation detection, classification and feedback of brain state within a real-time acquisition interval. The system provide a research base, which is stable, practical and extensible for the study of rtfMRI based BCI.2. For the inevitable outliers in the signal during the data collecting, which caused by abrupt head motion and equitment instability, a robust kalman filter method is proposed to detect the brain activation. Based on the sparsity of the noise, the method models the outliers using a sparse noise term, and estimates the magnitude of the sparse noise using real-time convex optimization methods, thus provides a way for real-time monitoring of data quality and real-time correction of the model parameters. Furthermore, it can suppress the influence of the sparse noise in real-time activation detection of brain regions. Experimental results show that the method can quickly and accurately estimate the occurrence time point and the magnitude of the sparse noise, and effectively improves the stability of the brain activation detection.3. BCI performs the brain state classification quickly and accurately. In this paper, a linear kernel support vector machine(SVM) is introduced to perform the online classification of brain states. For the decision value drifting caused by the voxel drifting when using linear kernel support vector machine to classify the fMRI signals. A support vector machine separating hyperplane online update method based on K-means cluster is proposed.The method uses K-means cluster algorithm to estimate the drifting center of decision values, and then achieve the online adaptation to the decision value drifting by online update the bias of the support vector machine’s classifier. A BCI system based on the three states classification of motor function is built. Experimental results show that the method is adaptive to the decision value drifting, and can significantly improve the accuracy of classifier in each level.