Image Segmentation Of Brain Tissue Based On Neural Network

Brain tissue segmentation of MR brain images is the key step for brain quantitative analysis. Compared with other medical image segmentation method, MR brain image segmentation is more challenging. This is mainly because, first, there is no clear border between different brain tissues (gray matter, white matter, cerebrospinal fluid), beside that, MR imaging process has more formation of artifacts than other medical imaging modalities (such as chemical displacement artifact, motion artifact, magnetic susceptibility artifacts, etc.), magnetic field inhomogeneity of the deviation may also affect the imaging. These factors make accurate segmentation of different brain tissue difficult. That is why a growing number of scholars get into the study of MR brain image segmentation.Fist made research on segmentation of magnetic resonance technology using neural network, and made Application of LEGION(Locally Excitatory Globally Inhibitory Neuronal Oscillator Network)to segment images of brain tissue. The results were compared with the traditional threshold segmentation. The brain images of pre-partition aims to exclude non-brain tissue and retain the structure of the brain. Then in this thesis, with two original algorithm improvements, the model performance analysis is made to the improved fuzzy self-organizing neural network model (IFKCN). On the basis of model, brain issue segmentation aims to part the white matter (WM),gray matter (GM),cerebrospinal fluid (CSF). In the process of partition, the key parameters of the network were studied.The main contributions of this thesis are mainly lie in the following aspects:1. Made research on MR brain image segmentation based on neural network. The local stimulate and overall reject oscillatory model was adopt to segment the brain issue, the results were compared with the traditional threshold segmentation.2. Pre-segmentation work. This paper proposed two brain boundary line extraction methods based on region growing and boundary tracking method. A comparison is mad between the two ones.3. Two improvements were made to the original fuzzy neural network: updated membership of adjustment and the introduced the adjustment factor. This paper proposed an improved fuzzy kohonen self-organizing neural network model (IFKCN). Performance analysis was made on the model.4. Applied the improved model for brain segmentation. Results were compared with the original one. In the segmentation process, research was made on two key paraments: the input pixel vector design and introduced the clustering validity function to assess.