基於等時性演化之區域單元競爭演算法

cell competition algorithm based on isochronal evolution

吳旻鴻(Min-Hung Wu)

指導教授 ： 陳中明

國立臺灣大學/工學院/醫學工程學研究所/碩士(2005年)

https://doi.org/10.6342/NTU.2005.00271

摘要

超音波影像的多目標分割，可以幫助臨床醫師分割出可疑的目標物，可以幫助檢測人員節省手繪目標物的時間，也可用於新進人員的訓練之用，並且在發展電腦輔助診斷的系統上，扮演著非常重要的角色。

但是，由於超音波影像並不容易識別，因為具有低對比、高雜訊、斑點、假影、穿透現象等等的問題。這些現象使得超音波影像的目標物邊界容易模糊而不易識別。為了能夠降低超音波本身物理特性的影響，而能夠找出腫瘤的邊界輪廓，並且能夠減少超音波腫瘤影像容易過度分割的情形，在本研究裡，提出了等時性演化為基礎的區域單元競爭演算法。

本研究提出的等時性演化之區域單元競爭演算法，可分為兩個主要步驟。首先利用影像濾波器，降低超音波影像的雜訊，並強化邊界的特徵。接著使用分水嶺轉換演算法，以過度分割的方式，找出所有可能的邊界，並產生初始的區域單元。在結束這樣的前置處理與分割之後，依據等時性演化規則，引導區域的成長與區域彼此間競爭的過程，而分割出影像之中的目標物。

每一個區域有其成長的速度，其速度由局部相似值與全域變異值所組成。局部相似值代表區域局部相似的程度，促使區域向外成長；全域變異值代表區域全域的相似程度，代表區域向外成長的阻力。

本研究在臨床超音波影像的實驗結果中，不僅能夠有效的分割出腫瘤目標物，並且利用時間微調的彈性，對於惡性腫瘤常存在的腫瘤目標物過度分割的情形，能夠有所改善，而且經由本研究的驗證方式，本研究方法分割腫瘤影像與專家手繪分割腫瘤影像的誤差小於不同專家之間的手繪誤差。

關鍵字

影像分割 ； 超音波影像 ； 區域單元競爭 ； 等位函數 ； 分水嶺轉換

並列摘要

Boundary extraction of multiple targets in a sonogram can help clinicians find out perceptible objects, save time for sinologists and be used for training. Moreover, it plays an essential role in developing computer-aided diagnosis systems.
Because of the intrinsic properties in ultrasonic images, such as low contrast, high noises, speckle, artifacts and so on, it is generally difficult to automatically identify the boundaries of the images. These intrinsic properties make the desired edges blurred and deteriorate the discriminability of the boundaries. In order to alleviate the influence of these physical problems and mitigate the over-segmentation of targets, we proposed a new segmentation algorithm, which was a cell competition algorithm based on isochronal evolution.
The proposed method composed of two steps. First, we used the image processing filters such as Gaussian filter and Sobel filter to smooth the images and enhance the boundary information. After that, we used the Watershed transformation to over-segment the images to capture all possible boundaries and generate the initial cells. After completion of pre-segmentation, cell competition was performed following the criteria of isochronal evolution, which were imposed on regions when they were growing and competing with one another.
Each region grew with its own viscosity which consisted of local similarity and global variation. Local similarity serves as the primary force for region expansion, whereas global variation plays the role of expansion hindrance.
The experimental results on the clinical ultrasound images showed that our algorithm could identify all objects of interest reasonably well. Moreover by adjusting the beginning time, we can alleviate the over-segmentation for malignant lesions. The algorithm was validated by comparing with manually outlined boundaries. The results showed that the differences between the boundaries derived by the proposed method and the hand-outlined boundaries were within the range of the differences among different observers.

並列關鍵字

Image Segmentation ； Ultrasound Image ； Cell Competition ； Level Set ； Isochronal evolution ； Watershed Transform

參考文獻

[22] 陳博量.“能量區域化與具中隔保留能力之區域單元競爭演算法,”台大醫工所碩士論文,2004.

[1] Adalsteinsson D and Sethian JA, “The fast construction of extension velocities in level set methods,” Journal of computational physics, vol. 148, pp. 2-22, 1999.

[2] Adams R and Bischof L , “Seed region growing,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 16, pp. 641-647, 1994.

[3] Bottigli U and Golosio B, “Feature extraction from mammographic images using fast marching methods,” Nuclear instruments and methods in physics research, vol. 487, pp.209-215, 2002.

[4] Chopp DL, “Computing minimal surface via level set curvature flow,” Journal of computational physics, vol.106, pp.77-91, 1993.