電腦輔助診斷應用在乳房超音波之研究

Studies on Applications of Computer-Aided Diagnosis to Breast Ultrasound

楊欣嘉(Hsin-Chia Yang)

指導教授 ： 李百祺

國立臺灣大學/電機資訊學院/電機工程學研究所/碩士(2007年)

https://doi.org/10.6342/NTU.2007.01450

摘要

乳癌不論在歐美或台灣都是女性癌症的主要死因，而唯有及早有效地辨別腫瘤的類別，才能早期治療。乳房超音波有著非侵入式、無游離輻射、不受乳房緻密程度限制、便於攜帶、即時等優點，但卻存在影像品質因人而異及判讀不客觀的問題。前者可以透過參數影像、可適性影像與高頻超音波等方法來改善；後者則希望以電腦輔助診斷，定量客觀地來輔助醫師或技術師判別。由於聲學特徵可以反映出組織的特性，可適性影像可以消除失真，高頻超音波可以提升空間解析度，所以若能將這些改善影像品質的方法和電腦輔助診斷相結合，應當能有效改善乳房超音波的診斷能力。因此我們欲透過有別於一般商用系統的實驗裝置來探討聲學特徵、高頻超音波和可適性影像對於分類表現的影響，希望了解這些直觀上能提升診斷成效的方法，是否確實有效，而改善程度又有多少？

本研究的電腦分析方法中，使用的聲學特徵為相對於背景的聲速及衰減係數，是由有限角度超音波電腦斷層掃描的方式重建而得。實驗裝置是用一個壓克力架子固定乳房在金屬反射板上，並以中心頻率為5.57 MHz的線性陣列探頭(Acuson L6/128, STI, State College, PA, USA)搭配數位陣列系統(digital phased array system, DiPhAS, Fraunhofer IBMT, St. Ingbert, Germany)來擷取數據。所使用的影像紋路特徵則有平均亮度，自相關係數，衍生自灰階共現矩陣及不可分離小波轉換的參數。所使用的影像型態特徵則為代表形狀的depth-to-width ratio和代表輪廓的normalized radial gradient。分類方法使用的是線性支援向量機，而評估成效的方式為leave-one-out與ROC分析。可適性影像使用的是generalized coherence factor (GCF)的方法，來改善組織不均勻性產生的誤差。高頻超音波系統使用的是單一聚焦的lithium niobate探頭(Onda Corporation, Sunnyvale, CA, USA)，發射訊號是中心頻率為25 MHz的啾聲(chirp)波形。

可能由於樣本數不足、訊雜比不夠高等因素，儘管聲學特徵、影像紋路和型態特徵之間的相關係數偏低；和聲學特徵的結合卻不全然能提升影像特徵的判讀，且無法單純由相關係數和各特徵原本的表現找到規則。可能可適性影像使對比解析度提升，主要線條可以突顯，所以和線條相關之特徵能有較好的分類能力。但也因此使原本紋路較均勻的良性影像變成類似惡性影像有著明顯的明暗變化差異，所以不利於有關均勻程度、明暗變化及亮度這些特徵之表現。可能由於高頻超音波的空間解析度提高，所以較有助於增進與均勻度相關之特徵的區分能力；但又因高頻影像中斑點雜訊較明顯，反而使和亂度、與明暗變化相關之特徵的區分能力下降。

總而言之，本研究評估了聲學特徵、高頻超音波和可適性影像應用在乳房超音波之電腦輔助診斷的表現。儘管直觀上，加入反應組織特性的參數或改善影像品質應當可以提升電腦分析的成效，但透過各種系統的測試和討論，卻發現這些方法並不是對所有特徵都有用，甚至也無法在整體的分類成效上看到顯著提升。所以有別於我們預期的，最後的結論為，聲學特徵、高頻超音波和可適性影像是不必然能提升電腦輔助診斷之能力。

關鍵字

電腦輔助診斷 ； 乳房超音波 ； 聲學特徵 ； 紋路分析 ； 有限角度超音波電腦斷層掃描 ； 高頻超音波 ； 可適性影像

並列摘要

Breast cancer is a leading cause of cancer death among women no matter in Europe, America or Taiwan. Thus, early diagnosis and treatment is necessary. Breast ultrasound beats other modalities since it is non-invasive, no ionizing radiation, able to use in dense breasts, and so on. However, it suffers from the problems of operator dependency and limited image contrast. The former can be reduced by computer-aided diagnosis (CAD), a technique using quantitative features to help interpretation. The latter can be solved by methods such as parametric imaging, adaptive imaging and high frequency ultrasound. It is known that acoustic features can reflect tissue properties, adaptive imaging can reduce distortion, and high frequency ultrasound can enhance spatial resolution. Therefore, if we can combine these methods with CAD, the diagnostic ability of breast ultrasound is supposed to be improved definitely. For this reason, we plan to investigate to what extent acoustic features, adaptive imaging and high frequency ultrasound can influence the performance of CAD through an unconventional experimental setup.

In this computerized method, the acoustic features were relative sound velocity and relative attenuation coefficient (to the background). They were reconstructed using limited-angle ultrasonic computed tomography with an imaging setup consisting of a digital phased array system (DiPhAS, Fraunhofer IBMT, St. Ingbert, Germany), a linear array (L6/128, STI, State College, PA, USA) with 128 channels, a center frequency of 5.57 MHz, and a metal plate reflector to acquire B-mode images, time-of-flight data, and attenuation data at the same time. The image texture features were autocorrelation coefficient, average brightness, and parameters derived from gray level concurrence matrix and from non-separable wavelet transform. The image morphological feature were depth-to-width ratio and normalized radial gradient. The dataset was trained and classified by a linear support vector machine, validated by leave-one-out method, and evaluated by the area under receiver operating characteristic (ROC) curve (Az). In adaptive imaging, a method termed generalized coherence factor (GCF) was used to reduce the focusing errors from tissue inhomogeneities. In high frequency imaging, chirp waveform with a center frequency of 25 MHz was transmitted and a lithium-niobate single-crystal focused transducer (Onda Corporation, Sunnyvale, CA, USA) was used.

Maybe because of the calculation errors, insufficient sample size, and the low signal to noise ratio, though the linear correlation coefficients (CC) between acoustic, morphological and texture features were relatively low, the combination of both of them for classification did not necessarily improve the discrimination ability in the CAD. Adaptive imaging was more benefit to the features related to linearity in discrimination maybe because of the improvement in contrast resolution so that the lines can be more obvious. However, this contrarily made the features related to homogeneity, brightness variation, and brightness perform worse in discrimination since formerly homogeneous benign images became significant in brightness variation as the characteristic of malignant images. High frequency ultrasound was more benefit to the features related to homogeneity in classification partly because of the high spatial resolution. However, maybe due to the more significant speckles in high frequency ultrasound, the features related to randomness and brightness variation performed worse in classification.

In summary, this thesis evaluates the performance of acoustic features, high frequency ultrasound and adaptive imaging on CAD. It turns out that these methods did not necessarily improve the discrimination ability, even though it has been taken for granted that they can make the diagnostic performances better.

並列關鍵字

computer-aided diagnosis (CAD) ； breast ultrasound ； acoustic feature ； texture analysis ； limited-angle ultrasonic computed tomography ； high frequency ultrasound ； adaptive imaging

參考文獻

[1] "General Information about Breast cancer,"

[2] "A snapshot of breast cancer,"

[4] M. L. Giger, "Computer-aided diagnosis of breast lesions in medical images," Computing in Science & Engineering, vol. 2, p. 39－45, Sep/Oct 2000.

[5] D. R. Chen, R. F. Chang, Y. L. Huang, Y. H. Chou, C. M. Tiu, and P. P. Tsai, "Texture analysis of breast tumors on sonograms," Semin. Ultrasound CT MR, vol. 21, p. 308－316, Aug 2000.

[6] C. M. Chen, Y. H. Chou, K. C. Han, G. S. Hung, C. M. Tiu, H. J. Chiou, and S. Y. Chiou, "Breast lesions on sonograms: computer-aided diagnosis with nearly setting-independent features and artificial neural networks," Radiology, vol. 226, p. 504－514, Feb 2003.