乳房腫瘤超音波影像之電腦輔助診斷： 
幾近不受系統對比度設定影響之影像特徵擷取

使用乳房超音波影像檢查可以顯著的減少昂貴且易引發病患緊張之不必要的組織切片檢查。然而，從乳房超音波影像中歸納出影像特徵來區分良惡性腫瘤的工作卻十分仰賴臨床醫師的判讀。不同的醫師在不同的系統參數設定下其所判讀出來的結果很可能會不同。雖然電腦輔助診斷技術可以減緩此一問題，然而綜觀過去乳房超音波影像電腦輔助診斷的研究卻有一共同問題，即其所常用的區域性影像特徵會隨著超音波影像系統參數的設定改變而改變。為減少此潛在問題，本研究提出Normalization的概念以減少系統參數設定之影響，更嘗試改變影像對比度來進行嚴格的實驗。本研究所提出的電腦輔助診斷系統以幾近不受系統對比度參數設定影響之影像特徵擷取，主要模擬臨床醫師診斷乳房腫瘤的模式來區別乳房超音波影像中的良性與惡性腫瘤，提高正確診斷率以達到早期治療的目的。有別於以往電腦輔助診斷乳房腫瘤之超音波影像的研究，主要在於醫師可依循其專業調整超音波影像系統對比度參數設定且不受限於超音波影像系統的機型，同時以乳房腫瘤外正常組織超音波影像的紋理當作參考指標找出Normalization之標準，優點是降低不同對比度參數設定造成的潛在差別，並正規化輔助診斷系統所轉換出的紋理特徵值。本研究的實驗材料是由台北榮民總醫院放射部周宜宏醫師所提供，包括91個良性腫瘤的乳房超音波影像和69個惡性腫瘤的乳房超音波影像。本研究以幾近不受對比度參數設定影響所擷取出的區域性影像特徵，分別為Cell-Based特徵、Contrast特徵、Variance特徵、Sliding-Window特徵與Fractal Dimension特徵等，並將影像特徵以Logistic Regression Function以及leave-one-out cross-validation的概念進行特徵選取與分類，選出最佳的特徵值組合，進而將腫瘤區別為良性或惡性。經實驗結果可獲致0.8824

關鍵字

特徵值；迴歸分析函數；分類；對比度獨立；超音波影像；電腦輔助診斷；乳房腫瘤；影像分析

並列摘要

Breast sonography is an effective approach to reducing the number of unnecessary biopsy, which is not only an expensive but also an unpleasant process. Nevertheless, to extract the image features for differentiating the malignant lesions from the benign ones is a task highly depending on the interpretations of medical doctors. Moreover, breast sonologists with different experiences or on different sonographic system settings might have different interpretations of the sonograms. Although these problems may be alleviated by taking advantage of computer-aided diagnosis (CAD) techniques, most previous CAD algorithms suffer a common problem that their textural features may vary with the system settings in general. To remedy this deficiency, we propose to minimize the influence of system setting by incorporating the idea of normalization. This idea is verified by a rigorous simulation study, in which the same images with different contrasts are simulated to test the robustness of the proposed CAD algorithm. Consequently, the uniqueness of the proposed CAD system lies in the nearly contrast-independent images features, which mimic the clinical strategy frequently used by medical doctors in differential diagnosis of benign and malignant breast lesions. Compared with the conventional CAD systems for breast lesions, the proposed CAD system has a salient feature that no constraint is imposed on the contrast settings, which allows the medical doctors to adjust the contrast freely to attain the best views of sonographic breast lesions. All the sonographic breast images are provided by the Department of Radiology, Division of Ultrasound, Taipei Veterans General Hospital, including 91 benign breast lesions and 69 malignant breast lesions. Several textural image features with nearly contrast-independent are proposed in this study, namely, Cell-Based features, Contrast feature, Variance feature, Sliding-Window features and Fractal dimension features. Feature selection and classification are accomplished by using a logistic regression function based on leave-one-out cross-validation strategy to select the best combinations of features for differentiating benign lesions from malignant ones. The experimental results show that the best performance achieved by using textural features only is 0.882±0.015 and 0.823±0.016 in terms of and classification accuracy, respectively. Although this performance is inferior to that achieved solely by using geometrical features, supported by the results of paired-sample t-tests, a better performance has been attained by combining both types of features, the best of which is 0.967±0.011 and 0.934±0.018 in terms of and classification accuracy, respectively. With the support of paired-sample t-tests, the results of the simulation studies on the effect of contrast variations show that there is no significant difference between the performance for the original images and that for the images with different contrast degradations. It suggests that the proposed sonographic features, including textural and geometrical features, be nearly contrast-independent. In conclusion, the proposed textural features are not only capable of complementing the performance of geometrical features, but also robust to the potential contrast variation in real practices.

並列關鍵字

Breast neoplasm, Sonograms, Computer Aided Diagnosis, Image analysis, Classification, Logistic Regression Function, Features, Contrast-Independent

參考文獻

[1] Pisani P, Parkin DM, Ferlay J. Estimates of the worldwide mortality from eighteen major cancers in 1985: Implications for prevention and projections of future burden. Int J Cancer 1993; 55: 891-903.

[3] Brown ML, Houn F, Sickles EA, Kessler LG. Screening mammography in community practice: positive predictive value of abnormal findings and yield of follow-up diagnostic procedures. AJR 1995; 165: 1373-1377.

[4] Zheng Y, Greenleaf JF, Gisvold JJ. Reduction of breast biospies with a modified self-organizing map. IEEE Trans Neural Networks 1997; 8: 1386-1396.

[5] Fornage BD, Sneige N, Faroux MJ, Andry E. Sonographic appearance and ultrasound-guided fine-needle aspiration biopsy of breast carcinomas smaller than 1cm. J Ultrasound Med 1990; 9: 559-568.

[6] Bassett LW, Ysrael M, Gold RH, Ysrael C. Usefulness of mammography and sonography in women < 35 yrs old. Radiolography 1991; 180:831.

國際替代計量

乳房腫瘤超音波影像之電腦輔助診斷：幾近不受系統對比度設定影響之影像特徵擷取

全文下載

主題瀏覽