Research On Medical Images Segmentation Method Based On Improved Markov Random Field Model

Medical image segmentation is a fundamental technique of medical image processing such as3D visualization and surgery navigator.In order to distinguish normal tissue and abnormal pathological changes, medical image segmentation is needed.It is a hot and challenging project in medical image processing.Firstly, a survey on image segmentation methods is given and the image segmentation method based on Markov Random Field (MRF) model is focused on in this thesis.The basic theory on MRF model and its application in image segmentation field is researched.Secondly, due to the complexity of medical images,combined with only one kind of image information, the traditional MRF model could not get ideal segmentation result. Therefore, an improved compound MRF model which integrates prior and boundary information of the image is proposed, a novel prior energy function which integrates pixel intensity and boundary information of the image in neighborhood system is constructed in this thesis.The compound MRF model can segment more precisely, especially for the image with weak fuzzy boundary and concave area. In order to improve the speed of segmentation, Simulated Annealing with Probability Table (SAP) algorithm is proposed to solve the compound MRF model for the first time.The SAP algorithm updates the label and boundary field based on probability tables and is applied to segment medical images successfully. The effectiveness and accuracy of the compound MRF model are proved through the experimental comparison and differences evaluation.Thirdly, the traditional MRF and the FMRF model are only defined on the certain class, they could not get ideal segmentation result for the images with wrapped fuzzy area and partial volume effect.Therefore, this thesis introduces the soft segmentation theory into MRF model and an improved FMRF model is established. The prior FMRF is updated based on maximum posteriori criteria and the segmentation result is obtained by defuzzifying the segmentation result. Experimental results show that the FMRF model can effectively segment images with fuzzy area and partial volume effect to ensure more accurate segmentation result. Finally, the summary and the prospect on this research are given.