Please use this identifier to cite or link to this item: http://hdl.handle.net/11455/6172
標題: 應用正交次空間投影法於多頻譜核磁共振造影腦部影像之分類
Orthogonal Subspace Projection (OSP) Approach to classifying Multispectral Magnetic Resonance Image (MRI) of the Brain
作者: 何效堅
Jian, Ho Hsiao
關鍵字: 核磁共振造影;Magnetic Resonance Imaging(MRI);特徵轉換;正交次空間投影法;特徵影像濾波法;抑制能量最小化法;類神經網路;目標點影像法;主成份分析;Feature transformations;Orthogonal Subspace Projection(OSP);Eigenimage filtering;Constrained Energy Minimization(CEM);Competitive Hopfield Neural Network(CHNN);Target point image(TPI);Principal component analysis(PCA)
出版社: 電機工程學系
摘要: 
在現今的放射診療學上,核磁共振造影(MRI)為目前應用最廣的造影術之一,核磁共振造影可以提供多方向的的解剖切面造影,使得疾病的診斷更加準確,以提高治療的正面效果。當一病人被安排做MRI影像後,對人體某一組織切面部位,會產生一系列的多頻譜影像。而隨著不同的參數 (如: TR, TE) 會產生包含不同訊息的影像 (如: T1WI, PDI, T2WI)。判讀人員可以利用不同影像的訊息來判定1.顯示腦的髓鞘發育2.判斷先天性腦部異常3.對缺血性及出血性中風的診斷4.老人老化的觀察5.顱內腫瘤6.腦部外傷之診斷7.中樞神經系統感染的診斷
8.腦部白質病變之診斷9.顱頸交接區病變的診斷。而同時使用多個頻譜,雖可獲得更多的資訊,但對判讀人員卻可能造成困擾。所以我們以下列討論之方法,將多頻譜所含的資訊經處理後以單一的影像來顯示,以提高判讀的效率。
我們的實驗是以正常的腦部影像為例,分別對CSF,白質,灰質來作強化,而其所用之方法亦可進一步的用於病變之強化。對於醫師的診斷能提供更有效的方法。因此本論文即在研究由多頻譜MRI影像中,把腦部的主要組織(如CSF [cerebrospinal fluid]、WM [White Matter]、GM [Gray Matter])分割出來。
在本研究中,我們先嘗試使用數種多頻譜特徵轉換各自對MRI影像做特徵擷取,這些轉換分別是主成份分析、目標點影像法、比值濾波器、類神經網路、抑制能量最小化法、特徵影像濾波法與正交次空間投影法。比較這七種特徵擷取的結果中,我們採用正交次空間投影法(Orthogonal Subspace Projection-OSP)當成主要的特徵擷取。關於正交次空間投影法運算中所須的兩個輸入因子─想得到的特徵向量(Desired feature vector)與不想得到的特徵向量(Undesired feature vector),本論文更進一步發展其相對應之自動選取方法。一組多頻譜影像經由正交次空間投影法後將可得到一張具有影像增強效果的特徵影像;亦即在此張特徵影像中,想得到的器官會顯現較亮的灰階值,不想得到的器官會顯現較暗的灰階值,如此一來將有助於後續的組織判別。
對於腦部MRI多頻譜影像,根據實驗結果顯示,本系統所提之方法均能成功的分割出CSF、WM、與GM等組織,而其中雖有少數幾張影像品質不是很好,但其餘均能有良好的結果,足以說明本研究方法是可行的。

In modern radiotherapy treatment planning, Magnetic Resonance Imaging (MRI) is one of the most widely used radiographic techniques. MRI provides three descriptions (coronal imaging, sagittal imaging and axial imaging) of internal structures to help doctors make accurate diagnosis of diseases. After a patient undergoes a MRI scan, a sequence of multispectral image slices is generated. Each slice represents one cross-section image of the three-dimensional human body. Following different parameters (ex. TR, TE) images will be generated which include different information (ex. T1WI, PDI T2WI). The different information derived from multispectral image will be utilized to diagnose (1).myelination of brain, (2).CNS congenital anomalys, (3).brain infarctions, (4).aging brain, (5).intracranial tumors, (6).head injurys, (7).CNS infections, (8).WM disease of the brain, (9).craniocervical junctions. Although using a multispectral image will get more information, it may create problems for the decision maker. So we further put the multispectral image through certain transformations to form one enhanced image which is more conductive to easy diagnostic decision making.
In our experiment we use a normal brain to illustrate and enhance the regions of CSF, WM, GM individually. These methods can further enhance the pathology of the brain to enable more efficient diagnosis. The goal of our work is to enhance the brain organs─CSF, WM and GM from MRI brain images.
In this thesis, six different multispectral feature transformations and one neural network approach - Principle component analysis (PCA), Target point image (TPI), Ratio filter, Competitive Hopfield Neural Network (CHNN), Constrained Energy Minimization (CEM), Eigenimage filtering and Orthogonal Subspace Projection (OSP) have been implemented and evaluated for feature extraction of MRI images. Among them, Orthogonal Subspace Projection shows its effectiveness, and is chosen as our feature transformation. The Orthogonal Subspace Projection needs two input factors - desired feature vector and undesired feature vector. We also propose a method to select these two feature vectors automatically. An image is obtained from five different spectral MRI images of the same cross-section of human brain and enhanced by Orthogonal Subspace Projection. In this enhanced image, the gray levels of the desired organ are brighter than those of the undesired organs. This property is useful for further diagnosis. The experimental results show that our proposed system is both effective and efficient.
URI: http://hdl.handle.net/11455/6172
Appears in Collections:電機工程學系所

Show full item record
 

Google ScholarTM

Check


Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.