| 摘要: | 一般而言，重建三維的醫學影像，通常會經由雜訊濾除、分割和三維內插等主要的步驟。本論文的重點是利用分割法和三維灰階值內插法於口腔核磁共振(magnetic resonance, MR)影像。 本論文的主要目的是利用改良的灰階值內插法由口腔核磁共振影像來重建三維的舌頭影像。我們首先利用蛇分割法(snake)來取得口腔MR影像內的舌頭組織，並提出了三種方法即形態灰階值內插法(morphology-based grey level interpolation，mb)、非修正侵蝕形態灰階值內插法(morphology-based non-correctional erosion grey level interpolation, mbnc)和修正的侵蝕灰階值內插法(morphology-based correctional grey level interpolation, mbc)來獲取新的內插影像。 我們的作法是先將兩張原始的二維影像提升到三維影像後重疊在一起，分出影像中物體的重疊，非重疊和背景區域後，再用膨脹距離轉換(dilation-based distance transform)，對這重疊後的三維影像距離轉換。然後對兩物體的重疊輪廓經由三種不同的侵蝕法而得到三種不同的輪廓，最後把三維的距離值轉為二維的灰階值便可完成影像外形和像素灰階值內插。另外，我們利用均方差值(mean squared difference，msd)，點位置不正確的數目(number of site of disagreement，nsd)，最大灰階值差異(largest difference of grey level，ldiff)，和正誤差(false positive)與負誤差(false negative)的總和即位置落差(location displacement，dispmt)等方法與與傳統的外形灰階內插和線性灰階內插法作比較。 雖然我們的結果(主要為mbnc)在msd (185.61)、nsd( 393638)和ldiff (120.67)等值比外形灰階值內插法稍微大(三值分別為181.75、 373661和 112.67)。但我們的結果具有最低的dispmt值，相較之下會有更平滑的外形。而此結果也符合我們重建三維舌頭MR影像及模型時，需要有較好的外形的要求。 General speaking, steps of noise removing, segmentation, and interpolation are often needed for the reconstruction of three-dimensional (3D) medical image. This study focuses on steps of segmentation and interpolation of our oral magnetic resonance (MR) images. The main purpose of this study is to reconstruct a 3D tongue from our oral MR images. We used snake algorithm to segment tongue tissues from oral MR images. Then we proposed morphology-based interpolation methods, namely, (1) morphology-based grey level interpolation method, (2) morphology based non-correctional erosion grey level interpolation, and (3) morphology-based correctional erosion grey level interpolation for the interpolated new images. Our method adopted grey-level interpolation to lift up 2D images to 3D images, and mark the overlap, non-overlap, and background regions. Then morphological inflation and erosion was adopted to interpolate the shape of objects (included mb, mbnc and mbc). After we finished distance code operation, the 3D images were collapsed to 2D images to complete image interpolation. In addition, we utilized mean squared difference (msd), number of site of disagreement (nsd), largest difference of grey level (ldiff), and the location displacement (dispmt) which is sum of false positive and false negative to compare our proposed methods with the linear grey level and shape-based grey level interpolation methods. Our results showed that mb methods have mixed performance in terms of values of msd, nsd, ldiff, and dispmt. Although the values of msd (185.61), nsd (393638), ldiff (120.67) of our method (mbnc) were slightly larger than those of the shape-based grey level interpolation (181.75, 373661, and 112.67, respectively), the objects which were interpolated by mbnc (which had the lowest dispmt value) had the best smoother contours among all methods. The smoothness of contours of the interpolated objects is the major concern of our application which is the construction of a tongue model that is based on the reconstruction tongue of our MRI data. |
