基於梯度向量流的適應性輪廓偵測應用在自然與醫學影像上之研究

林才槃
Snake, also known as active contour model, is widely used in searching for region of interesting (ROI) in different images. However, Snake has many weak points, and they do deeply affect the range of application, especially in the use of medical images. Many researchers tried to modify Snake algorithm to increase the performance, however, they were hardly to be used widely, because they were only used in specific kinds of medical image.

Other than the medical images, people also apply snake to natural images. However, the results from natural images were not successful because Snake is mainly based on gradient vector field to search for ROI, but the gray values of an image cannot efficiently contain color information. Therefore, it was not enough to search for ROI in natural images if we simply use Snake. This thesis takes the gradient vector flow field (GVF) as the base of the proposed system, and then it processes the medical images and natural images separately. The system is mainly divided into two parts. The first is the medical image. Because the histograms of medical images always flock in some specific area, the system uses the technique of image enhancement to raise the contrast of medical images to accentuate the boundary information of medical images efficiently. The second is the natural images. In order to present the color information of natural images efficiently, we firstly use Fuzzy c-means to segment the image, and then combine the results of the segmentations with the traditional gradient information. After using the above methods to deal with images, we use Canny to search for the edge map of the images to substitute the traditional gradient information of GVF. Comparing with the methods that people proposed before, which were mostly aimed at a specific medical image, this paper proposes a method that can be used in the three types of medical images, i.e., CT, MRI, and ultrasound images. As for the process of natural images, we add the information of color to supply the lack of GVF; therefore, we can find the outlines of our objects in the images and broaden the application of the GVF, which originally was used in medical images.

Snake或是稱做動態輪廓模型(active contour model),這各模型廣泛被運用在尋找不同影像的ROI方面。但是傳統的Snake模型有許多缺點,而這些缺點深深影響了Snake模型的應用範圍,特別是醫學影像方面的應用。此外有許多雖利用修改過的Snake模型來尋找醫學影像的ROI,然而這些系統都只適用於特定類別的醫學影像,所以並無法廣泛的運用。除了將Snake的模型運用在醫學影像之外,也有人利用Snake的模型來尋找彩色影像的ROI,但是可預期的是結果並不好,因為snake主要是透過影像的灰階向量場(gradient vector filed)作為尋找ROI的依據,但影像中的灰階值並無法有效的包含彩色的資訊,所以單純Snake 模型來尋找彩色影像的ROI是不夠的。本論文的系統是以修改過的Snake模型又稱為梯度向量流(gradient vector flow field )為基礎,再將醫學影像及彩色影像分開處理。本論文的系統主要被分為兩個部份。首先就醫學影像來說,由於醫學影像的灰階統計圖(histogram)都會集中在某個區域,所以本系統採用影像增強的技術(image enhancement)先提高醫學影像的對比度,這樣可以有效的突顯出醫學影像邊緣的資訊。至於彩色影像方面,我們先對彩色影像做初步的分割,本系統所採用的是 Fuzzy c-means 這個演算法對影像做分割,經過初步的分割後,將分割的結果與傳統的梯度資訊做結合,這樣即可有效的保存彩色影像的色彩資訊。在我們利用不同的方法突顯出不同影像的特性後,再利用Canny (Canny edge)來尋找影像的邊緣圖(edge map)來替代GVF的傳統的梯度資訊。本篇論文對於醫學影像的處理,分別在電腦斷層(CT)、磁核共振(MRI)以及超音波影像上都做了驗證,皆能準確找到ROI;而且對於有雜訊的影像,收斂的結果依然能找到ROI。比較前人所提出的方法,大部分只針對某一特定的醫學影像,本篇論文所提出的方法可以適應上述三種的醫學影像,不再受限於單一演算法只能處理單一的醫學影像。而對於一般自然影像的處理,因為加入色彩的資訊,彌補原本梯度向量流的不足,這樣便能在一般自然影像中更精確找到目標的輪廓;提升了原本只被廣泛應用在醫學影像的梯度向量流的應用範圍。