融合等位函數法與區域單元結構和圖形劃分資訊於乳房腫瘤超音波之分割

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。 本研究提出一種新的超音波影像分割技術，此法結合了等位函數法及區域單元競爭演算法之區域單元結構以及圖形劃分的資訊，應用於超音波的影像分割上面。由於超音波影像具有低對比、高雜訊、斑點、假影以及周圍組織相關紋理等現象，導致超音波影像的目標物邊界模糊而不易辨識。 基於區域單元競爭演算法可以將ROI區分出若干個我們有興趣的顯著區塊，例如；組織的結構、假影等。於是在本研究中一開始先利用區域單元競爭演算法得到影像中目標物最有可能的邊界，而後在等位函數法能量函數中加入區域單元邊界資訊的能量函數，將它視為目標物的形狀資訊，並利用Constrained Normalized Cut所得到的第二小的特徵向量去調整邊界資訊的重要程度。將整個等位函數法的能量函數中考慮到區域、邊界和目標物形狀的特性。利用較具視覺的區域單元邊界去吸引主動輪廓線，並克服微弱邊界和破碎邊緣的問題。 本研究利用了472張乳房腫瘤超音波影像作為驗證，其中包含221張惡性腫瘤，251張良性腫瘤影像，在本文最後的結果顯示，此方法可成功處理超音波影像目標物微弱邊界的偵測。

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. In this paper we propose a new method combining the level set method with the cell competition structure information and graph partition prior in ultrasound image segmentation. Because of the intrinsic properties in ultrasound images, such as low contrast, high noises, speckle, artifacts, tissue related textures and so on, it is difficult to identify the blurred boundaries or weak edges in ultrasound images. Since the cell competition algorithm can divide the ROI (Region of Interest) into several prominent components, which can be parts of the desired target, tissue structures, artifacts, and so on, we first apply it to capture the most likely visually perceived boundaries in ultrasound images. Then we regard the boundaries of the prominent components as shape prior knowledge and integrate it into the level set energy function. Next, we adjust the importance of the boundary information by a spatially variant band constructed from the second smallest eigenvector via the constrained normalized-Cut. The proposed method unifies the region-based, and boundary-based information as well as the shape prior in the level set energy function. The active contours are attracted by the visually perceived boundaries defined by the cell structures and the broken edges and weak edges of the cell structure are overcome by incorporating the second smallest eigenvector computed by the constrained normalized cut. The proposed algorithm has been validated on 472 breast sonograms which comprise 221 malignant and 251 benign cases. The result shows the proposed method can detect weak edges successfully in ultrasound images. Moreover, the boundaries derived by the proposed method are comparable to the manually delineated boundaries and robust in reproducibility.