電腦輔助診斷系統於乳房腫瘤X光與超音波影像之整合應用

Integration and Application of Computer-Aided Diagnosis for Lesions in X-ray Mammographics and Breast Ultrasonographics

賴宣宇(Hsuan-Yu Lai)

指導教授 ： 蘇振隆

中原大學/工學院/醫學工程研究所/碩士(2005年)

https://doi.org/10.6840/cycu200500815

摘要

本研究藉由對X光及乳房超音波之腫塊影像，進行辨別良、惡性之參數適用性評估及分析，以瞭解兩種類型影像之差異，並初步整合兩種影像類型之電腦輔助診斷系統介面。
在系統介面方面，我們設計一個操作方便且整合資料庫的應用之診斷介面；除了方便進行GVF snake輪廓點選的工作及影像處理參數可調控外，加上資料庫的管理，使得在參數取得更加方便外，更易於管理研究之資料。在輪廓圈選方面發現，雖然GVF snake輪廓圈選方法，可以有效的將輪廓參考點拉至邊界，但對於邊界資訊較複雜的影像，其輪廓圈選結果的再現性則受限於初始輪廓的位置。而腫塊影像之特徵參數皆具有描述形狀、邊界情形及影像灰階值分佈特性之功能；在進行這些參數對於良、惡性腫塊分類之適用性分析時，發現不同類型影像中，其參數之適用性會有所異同。接著，分別對X光及超音波影像腫塊之適用參數，進行類神經網路訓練；訓練結果發現，在X光影像類型中，關於描述腫塊形狀、邊緣及密度之相關參數，其重要性有相同的趨勢。而超音波影像類型中，關於腫塊形狀及邊緣相關之參數，其重要性較高。
最後，分別對10張X光影像(4張良性腫塊影像，6張惡性腫塊影像)進行類神經網路模型之訓練，並以另外17張影像(10張良性腫塊影像，7張惡性腫塊影像)進行診斷效果評估，發現accuracy為0.82、sensitivity為0.85、specificity為0.8及Kappa值0.63。而在超音波影像的診斷評估中，使用8張影像(4張良性腫塊影像，4張惡性腫塊影像)進行系統之訓練，而另外8張影像(4張良性腫塊影像，4張惡性腫塊影像)進行診斷效果之評估，則診斷效果之accuracy為0.875、sensitivity為0.75、specificity為1及Kappa值為0.74。在錯判案例的討論中，發現對X光影像而言，部分良性腫塊邊緣變化並非想像中的平順與近似圓形或卵圓形，加上類神經網路判讀之參數Circularity及l_Entropy的影響，因而容易產生誤判的情形。而超音波影像類型中，產生誤判的主要因素為GVF snake輪廓圈選的結果，易受雜訊的影響；再加上GVF snake輪廓圈選在不規則針刺狀的影像上，並無法忠實的呈現腫塊該有的輪廓，因而導致參數計算之誤差。
由P值與類神經網路訓練結果的分析中發現，無論對X光或超音波影像類型，Circularity都為重要的特徵參數；且所有特徵參數在不同影像類型中有不同的重要性，可進而瞭解系統診斷的規則。

關鍵字

GVF snake輪廓圈選 ； 類神經網路 ； 電腦輔助診斷 ； 超音波乳房攝影 ； X光乳房攝影

並列摘要

A computer-aided diagnosis system for masses in X-ray mammographics and breast ultrasonographics has been developed.We designed a thoughtful interface for users and discussed the differences of X-ray and ultrasound images with the characteristic parameters of masses.
We integrated an easy-to-operate interface and database into one system. That can help users finishing GVF snake contouring, image processing, and managing the data of each case. Some information was found in this study as follow: On the GVF snake contouring, we found that the re-appearance of contouring result could be limited from the initial position of the reference points on snake, especially to an image with complex edge information. To the parameters, all we used can well describe the shape, margin, density and size of each mass. From analyzing the characteristic parameters’ efficiency of dividing masses from benignant to malignant, we found some difference from X-ray and ultrasound images. On the training results of neural network, the order of importance of parameters are the same in X-ray images; the parameters of shape and margin are higher than others in ultrasound images.
Then we finished training and testing the neural network separately with X-ray and ultrasound images. In X-ray image, the number of training set is 10 (4 benignancy and 6 malignancy) and testing set is 17 (10 benignancy and 7 malignancy); the accuracy, sensitivity, specificity and Kappa value of CAD system are 0.82, 0.85, 0.8 and 0.63. In ultrasound image, the number of training set is 8 (4 benignancy and 4 malignancy) and testing set is 8 (4 benignancy and 4 malignancy). ; the accuracy, sensitivity, specificity and Kappa value of CAD system are 0.875, 0.75, 1 and 0.74. We also made some discussions on the cases of false diagnosis. To X-ray images, the reason of false diagnosis is that the margin and shape of some benignant masses are not so smooth or round and the influence of characteristic parameters-Circularity and l_Entropy is high. To ultrasound images, the reason of false diagnosis is the limitation of the GVF snake contouring, especially on the masses of irregular or lobular shape.
From analyzing of P values and neural network training results, we found that parameter Circularity is important to rules of system’s diagnosis in two types of images, and the order of importance for other parameters is different between X-ray images and ultrasound images.

並列關鍵字

X-ray Mammographic ； Breast Ultrasonographic ； Computer-aided Diagnosis ； Neural Network ； GVF Snake Contour

參考文獻

[4] A. Tuffanelli, S. Fabbri, A. Sarnelli, A. Taibi, and M. Gambaccini, “Evaluation of a dichromatic X-ray source for dual-energy imaging in mammography”, Nuclear Instruments and Methods in Physics Research, vol. 489, pp. 509-518, 2002.

[8] J. Schellenbach, and S. Steward, “Cancer facts & figures 2002”, American Cancer Society, pp. 9, 2002.

[10] 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, pp. 504-514, 2003.

[16] 陳逸雯，肺部腫瘤偵測之電腦輔助診斷系統，中原大學醫學工程研究所碩士論文，中壢，民國91年。

[17] Tam´as, Szir´anyi, and J. Zerubia, “Markov random field image segmentation using cellular neural network”, IEEE Transactions on Circuits and System, vol. 44, pp. 86-89, 1997.