利用等位函數之磁振造影影像皮質分割方法

Abstract

很多關於腦的研究選擇磁振造影影像來當作影像樣式，因為它具有較高的解析度以及在人體組織上有較好的鑑別率。而在很多腦結構及腦功能的研究中，大腦皮質(簡稱皮質)常是被關注的焦點。因此，從人的頭部磁振造影影像中將皮質分離出來是很重要也很基本的一個步驟。然而，在磁振造影影像中做大腦皮質分割時會遇到許多的難題，例如磁振造影影像的所產生的雜訊、亮度不一致、解析度不足以及人腦的複雜結構。 在我們的論文中，我們建立一個完整的皮質分割方法。首先，將影像中不屬於腦部組織的部分去除。接著我們使用一個混合的分割方法，以等位函數法為主再配合適應模糊C-均值聚類方法得到的資訊來做分割。適應模糊C-均值聚類方法可以解決影像亮度不平均的問題並計算影像中的點屬於各個腦部組織的機率。從適應模糊C-均值聚類方法得到的資訊可以用來計算等位函數法中的速度及初始三維表面模型。我們設計的區域運算元可以找出皮質的邊界而不被雜訊影響，而等位函數法中的三維表面模型最終將停在這個邊界。 我們的方法在具有雜訊及亮度不一致的影像中也可以得到不錯的結果。也可以適用於有不同腦部結構的影像。同時，速度也是另一個考量，在等位函數法及適應模糊C-均值聚類方法的計算中我們也加入了加速的方法，以求能更有效率的進行皮質分割。

Many brain researches use magnetic resonance imaging (MRI) as their image modality because of its high resolution and better discrimination of soft tissue. Because celebral anatomy and function analysis which are the major of brain researches focus on the celebral cortex, cortex segmentation becomes an important step. However, cortex segmentation from MR images is very problematic due to the di?culties such as imhomogeinities of the image, partial volume averaging (PVA) and the complicate structure of the cortex. In this thesis, we propose a complete segmentation system which are robust and e?cient. Level set method are used in our segmentation step. Adaptive Fuzzy C-means (AFCM) method are used to determine the fuzzy memberships of different tissues to estimate the initial surface and speed term which are important parts in level set segmentation. Alocal operator is applied to fuzzy memberships to produce a speed term which are robust against noise ofMRimages such as random noise, intensity inhomogeinities and PVA. An initail surface close to the boundary we want fo find is estimated. Techniques which can accelerate the AFCM and level set segmentation are applied in our method. Accoroding the techniques metioned above, we develop a cortex segmentation system. This system are robust to MRI of di?erent subjects. It is also robust to MRI with noise and intensitiy inhomogeinities. The segmentation system is efficient in each step because of the accelerating techniques.