Research On Modified FCM And Level Set Models Applying To Segment Infant Brain MR Images With Noise

Brain is one of the most important part of human body.Infancy is a critical period of brain development.However,the probability of suffering from a variety of encephalopathia in this period is quite high.Thus,early diagnosis method for infant brain disease has a great significance.Nowadays,in the medical diagnosis of encephalopathia,the main method is subjective observation and manual processing of brain medical images.Magnetic resonance image won't cause ionizing radiation to human body and has higher spatial resolution so that the imaging is the first choice in diagnosis of encephalopathia.Infant brain MR image not only is the shrink of the adult brain image,but also exists more serious noise,intensity inhomogeneity and partial volume effect,etc.These features cause that infant brain MR image processing is more difficult than adults'.One of the most important problems in the processing of brain images is the segmentation of brain tissues.My thesis aims to propose an improved algorithm to segment the infants brain MR images with noise.My research in this thesis lists as follow:(1)In this thesis,a modified Fuzzy C-Means(FCM)algorithm which is based on kernel function and local spatial information has been proposed.After studying the classical FCM model,I find that the method cannot accurately segment infant brain MR images with large noise.Through the study of several improved FCM algorithms not based on the gray value of single pixel,the Adaptive Spatial Fuzzy C-Means(ASFCM)has been chosen as my basic algorithm for further modification.In this thesis,the kernel function with nonlinear processing capability has been combined with the ASFCM algorithm to be a new algorithm named KASFCM.When solving the objective function of KASFCM,this thesis presents a novel iterative method,which does not require cross iteration membership function and clustering center and is advantageous to solve objective function,especially for some complex FCM models.Experimental results show that the KASFCM model proposed in this thesis has the ability to segment the infant brain MR images with large noise,which is superior to the FCM and ASFCM.(2)In order to further improve the ability of the brain MR image segmentation of infants with noise,this thesis proposes an improved ASFCM model based on non-local weight and kernel function(KNL-ASFCM).The idea of non-local weight takes full use of the information of more non-local regions in the image,not just the neighborhood information of some pixel point.Experiment results show that the KNL-ASFCM model has higher accuracy in infant brain MR images segmentation than other models mentioned above.(3)Aiming at the time-wasting issue of the above KNL-ASFCM model,this thesis proposes an method on the basis of FCM model and the level set model with modified mean information.To be specific,the local region information in the Local Binary Fitting(LBF)model has been fused into the Chen-Vase(CV)model as an improved CV model.Then,the improved CV two-phase model was successfully generalized to the four-phase model suitable for the brain MR images segmentation.Realizing the capability of this four-phase model to segment the infant brain images with large noise is weak,but the time consuming of the model is much less than KNL-ASFCM,this thesis therefore takes the KNL-ASFCM implementing for about 5 cycles or steps at first to segment the images and get an rough results aiming to provide the initial contours for the improved CV model to implement to get the accurate results.Experimental results through qualitative and quantitative comparison show that this novel method has the advantage of segmentation accuracy and time consuming in segmenting infant brain MR images with noise.