Study On Knee MRI Image Segmentation Based On Markov Random Field

Medical image segmentation is not only the basis of medical image analysis such as quantitative analysis and the three-dimensional reconstruction of the normal and diseased tissues , but the bottleneck of clinical applications. Segmentation accuracy is vital for doctors to assess the real situation of diseases and make correct diagnosis plans. However,due to the complexity,diversity,and degradation of medical images,medical image segmentation itself is ill-posed.A priori model based on Markov Random Field is a powerful method which is able to solve this ill-posed problem by employing priori information. It has been widely used for Bayesian image segmentation. In this dissertation,in order to segment knee Magnetic Resonance Imaging (MRI) images automatically ,accurately and quickly,we choose reasonable segmentation algorithm and strategy for particular tissues and get satisfactory results. The main work and contributions are as follows:(1) Bone segmentation in knee MRI can be regarded as the groundwork of segmenting and analyzing soft tissue in knees. Usually this task is time-consuming and needs human intervention. To solve this problem automatically and rapidly,a multi-scale MRF is introduced into knee MRI segmentation in this dissertation. Gaussian mixture model is firstly built as the statistical model for the intensity image,with an estimation of index number using MDL. In the phase of building multi-scale MRF model,non-iterated computing based on causality between scales is implemented,where statistical information is transferred from fine scales to coarse scales and MAP of every pixel is computed from coarse scales to fine scales. As a result,fast and unsupervised bone segmentation on knee MRI can be achieved. The experiments show that the temporal cost of segmenting knee bones based on multi-scale MRF is relatively low and the segmentation error can be similar to manual segmentation by medical experts. In conclusion,the work presented here accomplishes fast and accurate segmentation on knee MRI of low SNR through building a multi-scale MRF model. Future work can be extended to further cartilage and meniscus segmentation.(2) Magnetic resonance images have complex contents in both morphology and texture,which impose difficulty on effective image segmentation. To this end, the strategy of"segmenting-locating-extracting"is proposed,where MRI features and knee anatomical knowledge are made the most of as priori information. Firstly,multi-scale Markov Random Field method is used to implement an automatic and fast segmentation of tissues that have similar intensity distribution as menisci. Then the meniscus region is roughly located by combining Sobel operator with histogram projection. Finally,the areas of connective regions to extract the segmented meniscus anterior and posterior horns are determined accurately. Compared to related work on manually or semi-automatically segmenting menisci , the experiments show that the proposed algorithm automatically performs a repeatable and accurate segmentation on menisci,with relatively low temporal cost.