Reseach On The Segmentation Of Brain MR Image

The medical image processing technologies have had rapid applications in recent years. Magnetic Resonance Imaging(MRI) technology is a crucial visualization tool for clinical diagnosis. It is important in neuroscience for the identification of brain structures from MR image to provide quantitative assessment of medical image analysis, e.g., surgical planning, post-surgical assessment, abnormality detection, mapping of functional activation onto brain anatomy, the study of brain development, and the analysis of neuroanatomical variability in normal brains, and so on. Brain magnetic resonance(MR) image segmentation is one of the most important parts of clinical diagnostictools. Motivated by the importance of MR image segmentation, many researchers have paid much attention to partition the brain MR image into different regions based on given criteria for decades. However, the brain MR images mostly contain noise, inhomogeneity and sometimes deviation. Therefore, the accurate segmentation of brain MR images is a challenging problem for the relevant researchers.This dissertation presents work on segmentation of brain MR image, including two main aspects-the normal tissue segmentation and the white matter lesion segmentation, which is based on the energy minimization framework. The corelative researches cover the following parts.1.An automatic algorithm is proposed to simultaneously solve the tissue segmentation and bias field estimation.This dissertation focuses on the three dimensional segmentation of normal tissue(gray matter, white matter and cerebro-spinal fluid) from brain MR images. A salient advantage of this method is that its result is independent of initialization, which allows robust and fully automated application. In order to overcome the effect of noise, the nonlocal means technique is further introduced to achieve spatially regularized normal tissue segmentation.2.Extend the energy minimization framework into some conventional segmentation methods.Some conventional segmentation algorithms, e.g., k-NN, MRF, EM, are extended to combine with the energy minimization framework. These improvements overcome the limitation of misclassifications from above-mentioned conventional classifications.3.An efficient and robust algorithm is proposed for temporally consistent segmentation of longitudinal MR data acquired at different time points.The input serial images preprocessed by image registration are considered as a 4D image data. Our method jointly estimates the bias fields and obtains a segmentation result with temporal consistency. Temporal consistency in the segmentation results is achieved by regularizing a 4D vectorvalued function computed from the input 4D data and estimated bias field and the tissue intensity means.4.Propose the white matter lesion segmentation method from multichannel MR images.The white matter lesion segmentation is, thirdly, proposed to detect the white matter lesion from multi-channel images(T1-w, T2-w and FLAIR images). The formulation for normal tissue segmentation and bias field estimation in single channel image is extended to a multi-channel formulation. White matter lesions are considered as the fourth type of tissue, in addition to GM, WM, and CSF. In order to overcome the influence of noise and improve the accuracy of the final lesion segmentation, two improved methods are developed based on the multi-channel energy formulation. The nonlocal means technique is introduced in one improved method to achieve the spatially regularized lesion segmentation. In the other improved method, the above robust energy minimization approach is used to perform a preliminary segmentation of lesion, WM, GM and CSF, and simultaneously estimate the bias field to deal with the intensity inhomogeneities in the input MR images. In the second step, the lesion boundaries are refined by using a level set formulation that is able to exploit local intensity information to precisely locate the object boundaries. The preliminary segmentation provides a good initialization of the level set function for the level set method in the second step, which further improves the accuracy of the final lesion segmentation of the proposed algorithm.5.Propose an efficient and robust algorithm for temporally consistent of lesion from longitudinal multi-channel MR images.The proposed method can segment the small areas of dead cells(white matter lesion) and correct the bias field under an energy minimization framework for the multi-channel images in each time point. The nonlocal means algorithm is used to regularize the 4-D vector valued function computed from the 4-D data to achieve the temporal consistency for lesion membership function as the final segmentation result.At the end of this thesis, we conclude the advantages and disadvantages of the proposed methods, then give a guide as to the future work.