本研究的最終目標是利用建構口腔舌頭模型，來研究生理語音構音的機制。因此利用口腔磁振影像分割出舌頭構造所重建的舌頭模型來代表實際的舌頭大小是很重要的步驟，而舌頭磁振影像分割的好壞則關係著重建出來的三維影像準確度。本研究主要目的在節省時間人力成本考量下，對口腔磁振影像自動分割出舌頭構造，並與手動分割出來的結果做比較以瞭解自動分割的準確性。本研究提出結合等位函數法與梯度向量流蛇模型的方法，利用等位函數法做影像自動分割，再用梯度向量流蛇模型做修正舌頭邊緣並平滑化輪廓的步驟。結果顯示本研究一個個案分割的時間約需5.5分鐘，比熟悉舌頭構造操作者手動分割所需22.6分鐘來得快。本研究的準確評估方法是利用相似係數法，結果顯示我們用的方法對大部分切面之平均相似係數都大於0.88 (8位個案，4女、4男)，達到不錯的結果。比較本研究分割與手動分割重建後的三維影像在外形上大致相同只是有些不平整，但從中央矢狀切面來看舌頭內部構造在直覺上沒有差別。 The long term purpose of this research is to study the physiological articulation mechanism based on a three-dimensional (3D) tongue model that is reconstructed from oral magnetic resonance images (MRI). The accuracy of reconstructed 3D tongue depends on the results of image segmentation of tongue structure from oral MRI data. The main purpose of this study is to automatically segment tongue structure from the oral MRI data not only to save time and efforts for data processing but also to keep the accuracy of automatic segmentation the same as the manual segmentation. This study adopted Level Set (LS) method to segment image automatically and used Gradient Vector Flow Snake (GVFS) method to move the contours toward the tongue boundary and to smooth the segmented contours. The results of our study showed that 5.5 minutes were taken to segment one subject automatically. This is faster than the time needed (22.6 min.) for manual segmentation by a well-trained operator. Similarity index was used to evaluate the accuracy of our segmentation. The results by our method showed average slice similarity index is greater than 0.88 (8 subjects, 4 females, 4 males). This indicates excellent agreement. In addition, the 3D tongue reconstructed from this study is less smooth than by the manual segmentation, and the shape of the 3D tongue reconstructed from this study is approximately similar to the manual segmentation. Finally, the internal structure of the tongue observed from this study from the tongue mid-sagittal slice is visually the same as the manual segmentation.
