利用以影像細胞單元元素為主的二元主動尋找輪廓模型偵測超音波影像中的腫瘤邊界

Abstract

超音波影像由於其無侵害性，即時，方便以及便宜的優點已經成為現代醫學在初期的臨床診斷上一項非常重要的工具。但是也因為超音波影像具有高雜訊的特性，使得傳統上以像素為主（pixel-based）的主動尋找輪廓方法（snake model）在這裡並不適用。在這裡我們提出一種以影像細胞單元元素為主（cell-based）的二元主動尋找輪廓模型方法去對超音波影像進行影像分割。這種以影像細胞單元為主的方式可以解決一般主動尋找輪廓模型在局部極小值方面的缺點。我們利用多重解析（multi-scale）的高斯平滑化去對影像進行去除雜訊且可保留影像邊界的平滑，並利用分水嶺轉換去得到初始的影像細胞單元元素的資訊。根據影像細胞單元的概念，我們除了要制定能量函數外，也另外去建立了一種新的形變過程。另外，二元主動尋找輪廓中的內部與外部輪廓彼此之間也會提供一個拉扯的力去避免兩者之中任何一個被局部極小值所造成的陷阱牽制的缺點。最後，我們將會以電腦模擬的影像去驗證我們的結果，並實際應用在超音波影像的影像分割上。

Ultrasound imaging is an important diagnosis tool for early detection and regular check-ups because of its non-invasion, real-time, convenience and economy. Owing to the low signal-to-noise ratio in ultrasound images with speckle noises, it is difficult to segment with classical active contour methods, like the snake-balloon model. This study will propose a cell-based dual snake model to segment ultrasound images. The classical pixel-based snake model and its modifications are easily trapped at local minima and we propose cell-based approach to solve this problem. Initial cells contains the capture ranges near high measures of edges are produced by the multi-scale Guassian smoothing filters and watershed transform. The energy function is established based on the original intensities of cells and the movements of the active contour will be generated upon it. Dual snakes of inner and outer contours based on cells are also used to provide additional pull forces when one of them is trapped at local minima. This new approach will be tested on simulated and clinical ultrasound images and the performance is reported in this study.