Research On Segmentation Method Of Vertebrae Magnetic Resonance Images Based On Automated Hybrid Modeling

Abstract:In human society, spinal disease has become commonly and frequently, it influences people’s lives seriously. Spine magnetic resonance images play an important role in diagnosing and treating various spinal diseases. Accurate and efficient vertebrae segmentation is not only conducive to shorten the time consumed by the doctor diagnosed spinal disorders, but also can be used to assist in the treatment of diseases of the spine by calculating parameters of vertebrae. A person has33vertebrae. It will consume much time to segment the vertebrae by hand, meanwhile, the different operator also cause difference segmentation results. The MRI image computer quantization can improve the spine diagnostic accuracy and objectivity. Therefore, the thesis is dedicated to the research of suitable computer method to segment vertebrae from MRI images.In order to solve the above problems and improve the accuracy of segmentation result, the thesis has proposed a segmentation method of spinal MRI images based on automatic hybrid modeling. The method contains hybrid modeling and hybrid segment. The hybrid modeling process includes four parts, automatic statistical shape modeling based on Minimum description length, the local gradient modeling, the statistical texture modeling and hybrid modeling. The hybrid segment contains three steps:first, to make sure the initial position, second, to search the intermediate segmentation results in local gradient, third, to search the final results in hybrid model. The proposed method combines the advantages of ASM and AAM. So the segmentation results are more accurate in feature points and texture. In the segmentation process, the method considers the geometric characteristics of the vertebrae, so it can reduce the errors which due to an initial position close to the real silhouette.Spinal MRI image segmentation experimental platform is built in windows xp and matlab2011b environment, and it is tested. The platform is divided into four modules, get the training data, automatic modeling, segment and result analysis.25groups of expert segmentation results are used to build the hybrid model. The value of variance and the objective function of the generated shape model is less than the manual and function parameters method. After model has built,3groups of spinal MRI images are used as test data to test the proposed segmentation method, the result of the proposed method had a higher dice coefficient and lower over-segmentation and under-segmentation rate than AAM method. The experimental results show that the segmentation result by the proposed method is more accurate than traditional ASM and AAM. And the time of modeling and search are much less than traditional AAM. There are31figures,4tables and68references.