| Summary: | 碩士 === 國立中央大學 === 電機工程學系 === 102 === In this study, we applied image segmentation algorithm to automatically segment the tongue contour from the oral magnetic resonance images (MRI) in order to construct a three dimensional (3-D) tongue in real human size and to study the anatomical structures of tongue muscles and reconstruct these 2-D slice results into a 3-D tongue. Based on the suggestion of the previous study from our laboratory, we adopted shape prior and fuzzy clustering knowledge into level set algorithm for solving the problems of previous research. We enhanced the pixel contrast of the first slice of each subject with fuzzy clustering to let level set contour evolve easier. For each non-first slice, we calculated the initial contour from the segmented tongue contour of the previous slice, and the segmented tongue contour of the previous slice also worked as the shape prior energy term to improve the current contour evolution. After contour evolutions, we used gradient vector flow snake to smooth the contour, and achieved automatic segmentation of oral MRIs. We evaluated the results of this study with the ground truth of tongue with the similarity index, percentage of difference and root mean square error. The similarity index is more sensitive to the accuracy of the segmented results among other evaluation methods, and the average similarity index of 8 subjects was 0.898 which indicated the similarity of the segmented results of this study is quite promising when compared to the ground truth, and the shape of the reconstructed 3-D tongue is similar to the one segmented with manual approach. This study used fuzzy clustering could improve the segmented results of level set for the first slice of each subject and the segmented tongue contour of the previous slice as a shape prior term successfully, and also calculated initial contour automatically to enhance the result of original level set method.
|
|---|
