Research On Brain Image Segmentation Based On Spiking Neuron Network

The brain tissue segmentation of Magnetic Resonance Image(MRI) is crucial to the studying and diagnosing of brain diseases such as Alzheimer's, multiple sclerosis, epilepsy, and schizophrenia. The purpose of the segmentation is to assign distinct labels to different regions such as white-matter(WM), gray-matter(GM), and cerebrospinal fluid(CSF) in skull free brain tissue. When the MRI has light intensity gradient or noise, however, the existing multi-threshold segmentation method can hardly segment the WM, GM, and CSF into 3 regions.The light intensity gradient makes it difficult to distinguish foreground from background. Thus the multi-threshold segmentation method may fail to segment the MRI into 3 regions, as it doesn't take the spatial correlation of neighborhood pixels into account. In order to utilize this spatial correlation information, Spike Neural Network(SNN) is proposed to simulate pixels and their inter-connections, using the synchronous and asynchronous pulse times to split brain tissue into different regions. The SNN adopts accurate spike time-encoding to simulate the biological stimuli to a neuron, which is more computationally powerful than that of the traditional neural network using thresholds or activation functions like sigmoid.The accurate firing time is generated according to the pixel gray level through the Population Coding Method called Gaussian receptive field, then the double layer feedback network using delayed Izhikevich neuron is constructed based on the pixel spatial information and the spike time-encoding method. In the SNN processing layer, there are lateral connections between excitatory and inhibitory neurons(the ratio of excitatory and inhibitory neurons is 8:2). In the output layer, there are feedback full connections from output neurons to the processing layer, while the number of the output neurons is determined by the number of the segmentation regions.The synaptic delays are initialized according to the gradient information of the brain MRI and the weights are dynamically adjusted according to the learning rules combining STDP with winner-takes-all competition regulation. With the proceeding of dynamic weight adjusting, the spike times become more and more synchronous in the similar area of MRI and more and more asynchronous in the different area. Finally by decoding the last spike times of output neurons, the brain tissue masks are generated and this achieves the segmentation of 3 different regions of brain MRI.The prototype system is programmed by using clock-driven and event-driven jointed manner. The simulation experiments are conducted and compared with that of classical algorithms like K-means and Kohonen neural network. The experimental results prove that SNN based brain tissue segmentation is more accurate and more efficient than that of the traditional brain tissue segmentation methods.