能量區域化與具中隔保留能力之區域單元競爭演算法

Abstract

超音波影像之目標物自動邊緣偵測，可幫助臨床醫師找尋可疑的目標物，節省檢測人員手繪目標物的時間，還可以應用於超音波影像之電腦輔助診斷，新進醫師的輔助教學，對於醫療診斷與教學研究都有幫助。 但是，超音波訊號在物理上有些限制：低對比、高雜訊、斑點(speckle)、假影(artifacts) 、音波穿透現象以及週圍組織相關紋理。這些現象都會導致目標物邊界的模糊而不易識別，為了能找出腫瘤的輪廓邊緣與內部的中隔(septa)，避免目標物邊界的分割錯誤或過度分割等問題，我們提出一個以區域單元分割為基礎的能量區域化與中隔保留能力之區域單元競爭演算法。 我們所提出的方法可分為二個主要的步驟：前置分割步驟與區域單元競爭步驟。在前置分割中，利用各種影像濾波器去除雜訊，並且強化目標物的邊界，接著使用分水嶺演算法產生初始的區域單元。區域單元競爭步驟是以區域單元作為區域分割、合併的基本單位，藉由區域能量函數，引導區域單元的競爭過程。 區域能量函數考慮了局部區域、全域區域以及區域相鄰邊界。其競爭機制是以局部相似值保留區域單元間的中隔邊界，並取得局部區域之間的相似度，再利用全域變異值計算區域合併所需付出的代價。 在臨床超音波影像與假體實驗的結果中，我們不僅能夠找出最明顯的目標物，目標物內部的中隔邊界也可以完整的保留下來。此外，我們所提出的方法還擁有多目標物分割和高度彎曲邊界輪廓的能力，這些優點在傳統的主動尋找模型裡是很難做到的，例如：蛇動模型。這些結果也顯示了我們的方法於實際應用的可能性。

Automatic boundary extraction of multiple targets of interest in an ultrasound image can not only help clinicians find out most perceptible objects, but also save sonologists’ time for boundary delineation. Moreover, it is potentially helpful for the novice instruction and medical research. However, ultrasonic images suffer several physical problems, such as low contrast, high noises, speckle, artifacts, artifacts tissue related textures and so on. These problems either blur the desired boundaries or deteriorate the discriminability of the boundaries. In order to extract the boundaries of multiple targets of interest in a single image, and preserve the septa of objects simultaneously, we propose a new cell-based segmentation method called energy-localized and septa-preserving cell competition algorithm. The new approach consists of two main steps, namely, the pre-segmentation step and the cell competition step. In the pre-segmentation step, linear filters are used to remove various types of noises and enhance boundary information. After that, initial cells are generated by using Watershed transformation. In the cell competition step, cell competitions involving splits and merges are performed based on local energy minimization using cells as the operation units. The cost function is composed of the local similarity and global energy. While the former is associated with local energy between edge segment and its vicinity, the latter is defined by the global variance between two adjacent areas. Local similarity values are used to preserve the septa of targets and estimate local similarity between adjacent areas. Global variance values serve as the resistant force to merge. The experimental results on the clinical ultrasound images and phantom images show that not only have the most perceptible objects of interest been identified, but also the septa of object have been preserved. Moreover, the proposed cell competition has shown to be capable of identifying the highly winding contour, which is not easily captured by the conventional scheme, e.g., the deformable models. The results suggest practical feasibility of the proposed cell competition algorithm to identify multiple targets for ultrasound image segmentation.