醫學影像分割使用最短路徑演算法

Abstract

影像分割就是將影像中所需要的部分分離出來，我們將計算出包含這個區域所需的封閉輪廓線。在現今許多影像分割的演算法中，當影像中我們所要分割的區域並不是非常明顯時，所做出來分割的結果並不是很好。在以梯度為基礎(gradient based algorithms)的演算法中我們會得到不連續的邊，而以區域為基礎演算法(region based approaches)我們可能得到太多或太少的像素點(pixels)。在這兩種處理方法中，我們無法得到一個很好的封閉輪廓線將我們有興趣的區域標示出來。然而在近幾年提出的snake演算法中，對於一些形狀較為奇怪的無法做的很好，也較容易受到雜訊的影響。在本篇論文中，利用相似於snake的觀念，將真實的邊界給標示出來。在這個方法中我們給予一個真實的邊界所掉落的區域，利用計算幾何演算法及圖形演算法，找到一個封閉的輪廓線將最接近真實的邊界給標示出來。根據實驗結果在幾種型態的醫學影像中，我們都可以得到一些不錯的結果。

Image segmentation is to separate the desired region of interest from an image. We then need to calculate a closed contour to enclose the region of interest. Many of the segmentation algorithm cannot work well when the boundary of the region of the interest is not sharp. In this case, the gradient based algorithms could obtaine broken edges and the region based approaches could extract too much or too little pixels. In both cases, we cannot find the correct closed contour for the region of interest. The active contour approaches were designed to calculate the boundary directly. Unfortunately, depending on the designing philosophy, a snake is either hard to handle the concave case or very sensitive to noises. In this thesis, a method similar to the active contour approach is developed. The method needs to know the region where the true boundary locates. The computational geometry and graph algorithm are the employed to find a contour that best approximates the true boundary. Many experiments on different medical images were conducted. Most of them demonstrated good results.