Research On The Segmentation Algorithms Of Magnetic Resonance (MR) Brain Images With Intensity Inhomogeneity

The medical image segmentation technique is the key technology in medical image processing and analysis, which is one of the classical problems in medical image processing. The medical image segmentation is the precondition and basis for the understanding and explication of the medical image, and be of great importance in the clinical diagnosis, the pathological analysis and the treatment. It is commonly used in image analysis, registration, fusion, the measurement of anatomical structure and image reconstruction. Specifically, the results of image segmentation can be used to measure the volume of human organs, tissues or lesions. According to the quantitative measurement and analysis of the volume before and after the treatment, the doctor can predict, diagnose and develop the patient’s treatment plan; the results can be used on the reconstruction of3D data, visualization, the formulation and simulation of the surgical program; the results of the image segmentation can also be used for the remote medical expert system. In the case of network bandwidth resources are limited, the accurate image segmentation is critical for improving the local image classification based on the region of interest(ROI) and the progressive transmission speed. In order to obtain the quantitative information of the size and the appearance of brain lesions, and realize the three-dimensional reconstruction of brain structure, the segmentation of brain image is particularly critical.During medical image segmentation, noise, partial volume effects and intensity inhomogeneity are three mainly considerable difficulties. The existence of noise makes the segmentation region becomes discontinuous, directly affects the accuracy of segmentation results. The intensity inhomogeneity (also known as non-uniformity field), is a phenomenon of the brightness change slowly, in the same physiological organization or structure. It includes both shading artifacts and inherent non-uniform of tissue properties. And in the whole image, the bias field is continuous, smooth and slowly varying. It is mainly caused by the imperfection of the magnetic resonance imaging equipment, including the RF field, the static field and gradient field. The inherent non-uniform of tissue properties is that, there exist quite a few spatially different substructures with functions within each tissue class in the human brain. Due to the inherent regional differences in imaging-related properties across substructures, the intensities in different substructures, even in the same tissue class, are also more or less different. The imaging-related properties that cause the inherent intensity variation include the composition, density, and magnetic properties (spin-lattice relaxation time T1, spin-spin relaxation time T2) of different tissues at different positions.Although the affect of intensity inhomogeneities on MR images is not obvious on visually, because the human visual system can automatically correct the inhomogeneity. However, intensity inhomogeneities in MR images, which can change the local statistical characteristics of the image, and cause the brightness distribution of different physiological organizations overlap, and the segmentation of MR images is more difficulty than other image, are a major obstacle to any automatic methods for MR image.In order to improve the accuracy of the segmentation algorithm for Magnetic resonance brain image, we research the problem of the noise and intensity inhomogeneity in the brain image segmentation, in-depth research on the segmentation algorithm for Magnetic resonance brain images corrupted by intensity inhomogeneity. The main work of this paper is as follows:(1) Based on local image model, we propose a Secondary segmentation algorithm for Magnetic Resonance Brain image based on Local Entropy Minimization(SLEM), to overcome the impact of the intensity inhomogeneity. The tissue-based block method meets the local image model, which makes it is possible to overcome the impact of intensity inhomogeneity by using the idea of segmentation based on local region. Secondly, the use of information entropy theory realizes the optimization of segmentation region, makes the algorithm to keep the localized and find the minimum area of the region affected by the intensity inhomogeneity at the same time, so improve the accuracy and computing time of the algorithm. Then, for each segmentation region, the use of FCM algorithm is not only suitable for brain tissue images with fuzzy characteristics, but also run fast. Finally, the regional dynamic search of secondary segmentation algorithm, achieves the secondary segmentation for the misclassification pixels in the first segmentation result, further improve the accuracy of the algorithm.(2) We use the over-segmentation of watershed algorithm as an advantage to overcome the intensity inhomogeneity in the image, and proposed a segmentation algorithm based on Regional Dynamic Search (RDS) for MR brain images corrupted by intensity inhomogeneity. Regional dynamic search is established on the basis of two concepts of search window and segmentation region(also segmentation environment), which achieves the dynamic corresponding relationship between the window and region, not only improves the ability of the algorithm to overcome the intensity inhomogeneity, but also solves the problem of boundary effects appear in the existing segmentation algorithm based on region. We use the characteristics of over-segmentation region and tag matrix obtained by the watershed algorithm, to find the region segmentation environment for search window which satisfy the conditions continuously. Finally, the FCM algorithm is independently performed in the final segmentation regions corresponding to each search window, to determine the segmentation results of the pixels in the search window, thus completing the fast and accurate segmentation of the whole image.(3) We use the "denoising" operation on the image of segmentation result, and propose a secondary denoising algorithm for the segmentation of magnetic resonance brain tissue image with noise and intensity inhomogeneity. Firstly, in order to reduce the impact of noise, we use a half-soft threshold denoising algorithm based on wavelet to remove the noise in image. Secondly, the segmentation algorithm based on regional dynamic search for MR brain images proposed in the previous chapter, is used to reduce the impact of the intensity inhomogeneity for segmenting the image after denoising. Due to the influence of noise, there are significant misclassification points in the segmentation result, we call "noise points". Therefore, we proposed the secondary denoising algorithm for the segmented image, to segment the pixels which meet the conditions of noise point again. The experimental results show that when the noise level is high in the image, the secondary denoising algorithm improves the accuracy of the segmentation results significantly better than the denoising algorithm before segmentation.Above all, this thesis makes an in-depth study on the segmentation algorithm of Magnetic Resonance brain image, the main consideration is the impact of the noise and bias field on the segmentation results. The proposed methods overcome the impact of the noise and bias field, and realize the segmentation of MR brain images rapidly and accurately at the same time. The experiment is implemented on both simulated and real MR brain images, and compared with other published algorithm, prove the validity and accuracy of our proposed methods in terms of the accuracy and run time of the segmentation results.