Automatic Brain Tissue Segmentation Methods Based On Magnetic Resonance Head Images

In recent decades, medical imaging technology has made considerable progress, and appeared X-ray computed tomography (CT), digital subtraction (DSA), magnetic resonance imaging (MRI) and other imaging techniques. Comparing with other imaging techniques, MRI has no damage on the human body and high spatial resolution of soft tissue contrast, so it is becoming more and more popular in research and application, especially in complex anatomical structures of some tissues such as brain. MRI provides valuable reference to neuroscience physician, magnetic resonance imaging can show people a more sensitive anatomical structures and lesions, so we can directly observe the internal organization of the human brain lesions. To sum up, it can effectively improve the diagnosis accuracy of brain disease and develop effective treatment programs.Medical image processing technology made a great convenience for a doctor’s diagnosis, doctor can access to specific tissues of the shape, volume, multi-angle view through computer aided analysis. Medical image segmentation is an important part of medical image processing, and it is the base of follow-up registration, reconstruction and quantitative analysis. This dissertation focuses on automatic brain tissue segmentation based on magenetic resonance head images, and the contributions of this dissertation are as follows:(1) An edge protected synchronization algorithm of MRI denoising and enhancement is proposed. There are some characteristics of brain MR images with edge blur, containing noise and poor contrast. They are seriously disturbing image segmentation and registration accuracy. The traditional approach is used to sovle the problem through enhancing after denoising or denoising after enhancing method, but the results are usual not ideal due to the contradiction inherents in this approach. The chapter proposes a new model which can protect the edge of image, and take care of denoising and enhancement. By combining with the adaptive histogram partial differential equations, we choose the PM algorithm as a basis and add the adjustment items to complete the synchronization of image denoising and enhancement. The results of experiment show that this algorithm has accuracy and practicality.(2) Through improving CV model in three fields, the chapter presents a new method for brain tissue extraction. Firstly, segmentation speed is enhanced through improving the classical distance function, it can accelerates the distance function convergence more rapidly and not reduces the accuracy simultaneously. Secondly, the improved method can change the uniqueness of classical result. The evolving lines will be stopped at the same level gray, so the spinal fluid can be wiped off. White matter and gray matter will be extracted more accurately. Finally, we present a dynamic condition for ending iteration through comparing the interval frame. This improvement changes the flaw of setting evolving times to end iteration, it can make the veracity and speed better. The methods are generally applied to2D and3D image segmentation, and the results of experiment indicate these improvements can make the brain tissue extraction more rapid and accurate.(3) Organizational structure of skull base is complex, some non-brain tissues are often associated with brain tissues, and the gray of them is very similar, most segmentation algorithms are difficult to complete the part of brain tissue extraction. For the organization "glue" problem of skull base images, a method of CV model combining with the morphological method is raised to segment these images. The CV model is firstly used to segment the input image. Because some tissues’gray is the same with the gray of eyes and other organs, but CV model is based on the regional average grayscale, so it can’t extract brain tissue accurately. To solve the problem, a morphology method of corrosion expansion algorithm is raised. It corrodes the border to make the adhesion organization separated, and then expands the border to the previous position to get the accurate target.(4) CV model is a method of extracting single object, to be able to simultaneously extract multi-object such as gray matter and white matter brain tissue, a method of brain tissue segmentation based on the multi-objective CV model is raised. Firstly, improved k-means clustering algorithm is used for rough segmentation to avoid the local optimal solution of multi-objective CV model. In order to improve the speed of k-means clustering convergence, the initial selection is also improved. The precise segmentation is done by means of improving the Heviside function through improved multi-objective CV model. The validity and feasibility of the method is verified by the experiments.(5) To make the results of manually segmentation more convenient and accurate, an interactive tumor segmentation method is put forward with improved Live-Wire model and NGGVF Snake model. Firstly, we manually draw the tumor’s contours by improved Live-Wire model to get the initial contours. Advantage of the improved model is that we can select less interaction points to get more accurate contour, especially in the larger part of curvature. Secondly, the initial contour is modified through NGGVF model. It not only overcomes the problem of local optimal solution, but also deeps groove part of the curve and accelerates iteration so it can improve the accuracy of the segmentation. Experimental results demonstrate the accuracy and validity of this method.(6) As the methods of brain tumor segmentation are often semi-interactive, disadvantage of these methods are not suitbale for bulk data processing, so an automatic method of brain tumor segmentation is proposed based on clustering tree matching for large number of brain MR images. Firstly, we locate the center of the segmented brain images and establish a matching structure of index tree. Coarse segmentation is completed using a node matching algorithm. Secondly, according to the coarse result of segmentation, precision division is adopted from deformation model or dynamic threshold region growing method. The algorithm can automatic complete the segmentation of brain tumors more accurate, and the practicality and feasibility of the algorithm are verified by experimental results.The dissertation designs and implements some relevant brain tissue segmentation algorithms in three directions:synchronization algorithm of MRI denoising and enhancement, conventional brain tissue segmentation algorithm and lesions (tumor) brain tissue segmentation algorithm. The results of denoising and enhancement algorithm can be used as input data for segmentation, and the results of segmentation are the subsequent processing basic steps of tissue measurement, registration, lesions determination etc. So these algorithms can support computer aided diagnosis system, and they can also provide the necessary help to diagnose brain diseases for doctor.