透過您的圖書館登入
IP:3.142.210.226
  • 學位論文

疊代式特徵映射聚集之超音波影像分割演算法

Iterative Eigenmap Aggregation for Boundary Delineation on Ultrasound Images

指導教授 : 陳中明

摘要


由於超音波沒有放射線的顧慮,而且檢查費用較低及操作方便。因此,醫學超音波乳癌影像診斷成為臨床上醫生用來篩檢乳房腫瘤的有效工具之ㄧ。超音波影像中的腫瘤形狀是醫師辨別腫瘤良惡性的重要指標,若能從超音波影像正確描繪出腫瘤的輪廓,並以電腦輔助診斷(Computer-Aided Diagnosis)技術提出建議性診斷,將有助於提高醫生判斷腫瘤良惡性的正確率。雖然手繪能夠勾勒出腫瘤形狀與輪廓,然而非常費時,也因此發展半自動或全自動影像分割技術來獲取腫瘤輪廓在實用上是必需的。 超音波影像分割是一個非常具有挑戰性的研究領域,因超音波影像具有高雜訊(high noise)、低對比(low contrast)、斑塊(speckle)、假影(artifact)以及周圍組織相關紋理等現象,使得超音波影像內部結構複雜而導致區域內部灰階值分布不一致。本研究之目的在於將參考人類的視覺訊息,透過結合影像結構、先驗知識與物件型態有系統的描繪出超音波影像中目標物之輪廓,提供電腦輔助診斷系統中前置處理所需之腫瘤形狀的輪廓,以便後續特徵擷取分析,成為醫生判斷腫瘤良惡性之重要依據。 本論文提出一個整合影像結構、先驗知識與物件型態之影像分割演算法,目的在於有效擷取超音波影像中,乳房腫瘤邊緣之輪廓,主要分成兩個階段。首先,對有興趣之目標物區域透過特徵值分解取得限制正規化劃分之第二特徵向量。接著,根據此特徵映射圖定義包含邊界與區域之能量方程式。最後,利用疊代式圖形劃分可以很精確地找到目標物之邊界。 在效能評估與分析中,將本演算法所產生的腫瘤輪廓與醫生手繪的腫瘤輪廓作比較,主要測試110張乳房超音波影像,分別包含60張良性腫瘤與50張惡性腫瘤。結果顯示本演算法所描繪出的腫瘤邊緣超過75%落在醫生手繪的腫瘤邊緣內,且威廉指標的值為1.085代表本演算法所產生的腫瘤輪廓與醫生手繪的腫瘤輪廓具有高度的相似性。另外本演算法所產生的腫瘤輪廓與醫生手繪的腫瘤輪廓之重疊面積比例大於0.90且其面積差異的比例小於0.14。由此可知本演算法適用於各種不同類型的乳房超音波影像腫瘤之擷取,且與醫生手繪的腫瘤輪廓具有一樣的參考價值。

並列摘要


Breast sonography is one of the effective tools for identification of breast cancers. It has been widely used for diagnosis and screening because of its non-radiation, low cost and effectiveness. The lesion shape in a breast sonogram is a crucial indicator for differentiation of benign and malignant lesions. If a computer-aided diagnosis (CAD) system is able to delineate the lesion contour delineated accurately and proposes suggestive diagnoses, it would be of great help for medical doctors in distinguishing the benign breast lesions from malignant ones. Although the lesion shapes may be demarcated manually, it is usually to time-consuming to be practical. Therefore, it is compelling to develop semi-automatic or automatic segmentation algorithms to acquire lesion boundaries in practice. Ultrasound image segmentation is a highly challenging task because of the high noise, low contrast, speckle, artifact, and surrounding tissue textures commonly found in a sonogram. These phenomena constitute a complicated image structure and result in inhomogeneous spatial distribution of gray levels. This study aims to delineate the boundary of the object of interest in an ultrasound image by referencing perceptual grouping in human vision and combining the image structure, prior knowledge and object morphology. The derived lesion boundaries can then be used as the basis of a CAD system, generating the features and suggestive diagnoses for the reference of medical doctors. This thesis proposes a new algorithm for boundary delineation of sonographic breast lesions, which integrates the image structure, prior knowledge and object morphology. The proposed algorithm is composed of two major steps. In the first step, the second eigen-vector map of the region of interest (ROI) is obtained by solving the constrained eigenvalue problem using eigen value decomposition. The energy function is constructed incorporating both boundary and regional properties of the ROI. In the second step, the lesion boundary is identified by an iterative graph cut approach. To evaluate the proposed algorithm, the lesion boundaries derived by using the proposed algorithm and those manually delineated boundaries by four expert observers on 110 breast sonograms have been compared, including 60 benign lesions and 50 malignant lesions. The results show that >75% of the derived boundaries lie within the span of the manually delineated boundaries. The Williams Index is 1.085 indicating that the derived boundaries agree as much with the manually delineated boundaries as the manually delineated boundaries agree with one another. The overlapping and difference ratios between the derived boundaries and the average manually delineated boundaries are mostly higher than 0.90 and lower than 0.14, respectively. The performance figures show that the propose algorithm is capable of deriving the lesion boundaries that are comparable to those demarcated manually, even for those breast sonograms with weak edges and artifacts.

參考文獻


[1] Chang RF, Wu WJ, Moon WK, Chen DR, “Automatic ultrasound segmentation and morphology based diagnosis of solid breast tumors,” Breast cancer research and treatment, vol. 89, no. 2, pp. 179–185, 2005.
[2] Chen, C.-M., Lu, H. H.-S., and Lin, Y.-C, “An Early Vision Based Snake Model for Ultrasound Image Segmentation,” Ultrasound in Medicine and Biology, vol. 26, no 2, pp. 273-285, 2000.
[3] M. Kass, A. Witkin, and D. Terzopoulos, “Snakes: Active Contour Models,” International Journal of Computer Vision, vol. 1, no. 4, pp. 321-331, 1987.
[4] Chenyang Xu and Jerry L. Prince, “Gradient Vector Flow:A New External Force for Snakes”, in IEEE Proc. Conf. on Computer Vision and Pattern Recognition, 1997, pp.66-71.
[5] N. Archip, R. Rohling, P. Cooperberg and H. Tahmasebpour, “Ultrasound image segmentation using spectral clustering,” Ultrasound in Medicine and Biology, vol. 31, no. 11, pp. 1485-1497, 2005

延伸閱讀