- 學位論文
以邊緣特徵為基礎之醫學影像切割
Medical Images Segmentation Based on Edge Features
摘要
本論文主要是發展一套以邊緣特徵(edge features)為基礎之醫學影像(medical images)偵測(detection)與切割(segmentation)技術，分別從子宮頸細胞(cervical cell)影像切割出細胞質(cytoplasm)、細胞核(cell nucleus)以及二維電泳(2-D electrophoresis, 2D gel)影像中切割出蛋白質點(protein spots)的輪廓，以便未來能正確分析子宮頸細胞正常(normal)、異常(abnormal)與蛋白質(protein)相關資訊。 生醫影像處理(biomedical image processing)為臨床病理學上最常使用的輔助工具之一；醫生在診斷任何病症時，藉由電腦的輔助進行自動判讀，減少人為疏忽，以提高判讀之效能。在本論文中，我們將分別以子宮頸細胞(cervical cell)與二維電泳(2-D electrophoresis, 2D gel)兩種醫學影像，做為本論文醫學影像(medical images)偵測(detection)與切割(segmentation)的主要研究分析對象。這兩種醫學影像最大的不同為：子宮頸細胞影像是以高倍率影像且單一個細胞質內包含細胞核影像；二維電泳影像則是多個顏色與形狀不一的蛋白質點影像。由於影像特性與使用者取材的對象不同，因此本文亦針對不同特性做不同的偵測與切割處理。 子宮頸細胞影像的偵測與切割技術，首先採用平均濾波器(mean filter)技術，有效的去除影像中的雜訊(noise)。再提出一個新的邊緣加強(edge enhancement)技術，係利用每一個像素(pixel)的粗糙(coarse)程度，做為加強物件邊緣的依據，再配合鄰近的二群法(two group)之邊緣梯度(gradient)加強技術以加強物件邊緣。然後運用Sobel邊緣偵測法(Sobel edge detection)偵測出物件邊緣資訊，其步驟包括：計算整張影像的梯度(gradient)，細線化(thining)是以非最大值刪除(non-maximum suppression)技術將邊緣以線形方式顯示，並配合遲滯性門檻(hysteresis thresholding)來決定可能物件的邊緣輪廓，利用HMTS演算法來取得邊緣的骨骸，以及採用分水嶺(Watershed)式的區域切割法，來決定出各邊緣線是否屬於簡單封閉曲線的物件。最後，定義子宮頸細胞影像上兩類邊緣線段特徵：一為兩條輪廓的邊緣線長度；二是輪廓的邊緣線為簡單封閉曲線，作為決定子宮頸細胞輪廓的方式。本文的實驗結果分別與Otsu’s以及等位函數法(Level Set)做子宮頸細胞抹片影像細胞質與細胞核的輪廓切割比較，除了將各種方法所得細胞質與細胞核的輪廓切割結果做比較之外，也做了影像切割的效能比較，實驗結果顯示本文方法能有效且精確地提供子宮頸細胞影像的切割。 二維電泳影像的偵測與切割技術，主要係利用切割出蛋白質點來決定其座標位置、大小、形狀以及濃度等資料，以分析二維電泳影像的一個基本動作。本文以子宮頸細胞影像的偵測技術為基礎，在取得二維電泳初始的邊緣梯度影像後，再將梯度假設以 函數分佈做為加強梯度之權重，以加強物件輪廓邊緣線，並將其鄰近像素間的邊緣視為各自獨立的輪廓。為了有效將輪廓做一完整的取得，以做為切割蛋白質點之依據，我們將影像中的輪廓分類為五種輪廓，分別是：簡單封閉輪廓(Simple Close Contour)、近似封閉輪廓(Approximate Close Contour)、破碎輪廓(Fracture Contour)、鄰近輪廓(Neighbor Contour)、以及非點對點近似封閉輪廓(Approximate Contour Determination for Not Among Two Extreme Points)，並分別做蛋白質輪廓過濾(spot filter contour)處理，其方法包含：近似輪廓偵測(Approximate Contour Dtermination)、破碎輪廓修補(Mend Fracture Contour)、鄰近輪廓處理(Neighbor Contour Processing)以及非點對點近似封閉輪廓偵測(Approximate Contour Determination for Not Among Two Extreme Points)等輪廓辨識。本文的實驗結果在邊緣偵測部分分別與Sobel、Canny、Roberts、Prewitt、Laplacian of Gaussian與Zero-Cross等六種傳統的邊緣偵測方法進行輪廓比較與分析差異。其中，在輪廓偵測部份則與等位函數法比較其結果差異。另外，為了強調本方法的適用性，將與商業軟體ImageMaster比較輪廓切割的成效，並從實驗結果顯示本文提出的方法能夠精確的切割二維電泳影像。
並列摘要
The main purpose of this thesis is to develop contour detection and segmentation techniques based on edge features for medical images. In order to accurately analyze the information on normal/abnormal cervical cells and protein, the proposed methods segment cytoplasm and cell nucleus of a cervical cell image and protein spots contour of a 2-D electrophoresis (2D gel) image, respectively. Biomedical image processing is one of the most frequently techniques in clinical diagnosis. Computer aided reading and judgment reduces manual ignorance and improves the judgment performance when doctors diagnosis. In this thesis, cervical cell images and 2-D electrophoresis (2D gel) images are the major research subjects to be analyzed. The difference between cervical cell images and 2-D electrophoresis images is the cervical cell image has a single cell within cytoplasm and nucleus in high image quality; the 2-D electrophoresis image contains lots of proteins in various colors and shapes. Due to different image characteristics, this thesis proposes different kinds of object detection and segmentation techniques for suiting to cervical cell images and 2-D gel images. For the detection and segmentation of the cervical cell image, the mean filter is used to remove the noise of the cervical cell image. Then, an edge enhancement technique using the coarseness of each pixel and the gradient of the nearby two-group method for enhancing object edges is proposed. Then, Sobel edge detection is used to detect the edge information of the objects: Calculate the gradient of the cervical image and use thinning to express the edges in linear shape with non-maximum suppression. Then, Hysteresis thresholding is used to determine the possible object edge contour, and HTMS is applied to obtain edge skeleton. At last, watershed regional segmentation is used to determine whether edges belong to closed curve objects. Finally, the characteristics of cervical cell image edge lines are defined: The length of the edge lines of the two contours and edge lines of the contour being simple curve to determine the methods of cervical cell smear images. The experimental results are compared with that of Otsu and level set as cervical cell smear image periblast and cell nucleus contour segmentation. Also, the performance of mage segmentation is compared. The results show that the proposed method of this thesis can effectively and precisely offer cervical cell image segmentation. The 2D gel image detection and segmentation are preprocesses which segment the protein points to determine their coordinate locations, sizes, shapes and density before diagnosis analyzing. Based on the cervical cell image detection technology, the initial edge image of 2D gel can be obtained by above techniques, and logarithm function is applied on the gradient of the initial edge image for edge enhancement. In order to effectively obtain the complete edges for segmenting the protein points, the contours of the edge image are classified to five different kinds of contours: Simple close contour, approximate close contour, fracture contour, neighbor contour and approximate contour determination for not among two extreme points. At last, above contours are further dealt with spot filter contour for obtaining compete contours, the spot filter contour including approximate contour determination, mend fracture contour, neighbor contour processing and approximate contour determination for not among two extreme points. From the experimental results, the edge detection of the proposed technique is compared with the six traditional methods which are Sobel, Canny, Roberts, Prewitt, Laplacian of Gaussian and Zero-cross for deviation analysis. Besides, the proposed contour determination method is also compared with level set method. Moreover, to emphasize the adaptability of this method, the proposed method compared with the software “ImageMaster” on contour segmentation for the performance. Finally, the experimental results show that the proposed method could precisely segments the 2D gel images.