應用三維型態分析的乳房MRI電腦輔助診斷

Abstract

過去數年來，乳癌一直是全球女性主要的死因，但是其治癒的可能性，也隨著早期發現以及完善的醫治而提高，美國癌症協會也建議婦女需每年定期作乳房檢查，以便早期發現可疑的腫塊；而截至目前為止，電腦輔助診斷系統已經不只可以單純提供所選定腫瘤的資訊，更可以藉此區分腫瘤的良惡性，因此，對於腫瘤的切片檢查次數就可以相對的減低。一般來說，腫瘤的型態可以反映出腫瘤的良惡性，因此在這篇論文內，使用了三維高解析度的核磁共振攝影影像來進行腫瘤的診斷，我們使用數種有關型態的特徵值，來描述所選定的腫瘤，更進一步地，我們亦利用腫瘤的形狀建立一個三維的橢球模型，藉由比較此模型與腫瘤的異同點，我們亦定義了許多描述腫瘤型態的特徵值；除此之外，我們亦利用灰階值共生矩陣對腫瘤進行紋理的分析，並把其和之前提及的腫瘤型態分析互相比較。在我們的實驗裡，我們對總計包含95個經過病理學驗證的腫瘤，重複以不同的特徵值種類測試，其中包含44個良性病例以及51個惡形病例，根據實驗結果，我們發現與腫瘤型態相關的特徵值比起紋理的特徵值，更能區分出腫瘤的良惡性，而我們最後討論選定的數種特徵值，能達到準確性88.42%、敏感性88.24%以及特異性88.64%。

In the past years, the breast cancer is globally the major cause of the death for women. But the curability of the breast cancer can be raised if such a tumor can be found early and treat correctly. American Cancer Society has also suggested that women should take the breast examination annually to achieve the early detection of the suspicious tumors. Recently the computer-aided diagnosis systems can help radiologists not only to render the information of the tumor but also to different the malignant tumors from benign ones. Hence the demand of the breast biopsy of the found tumors can be further reduced. In this paper, the three dimensional (3-D) high resolution magnetic resonance imaging (MRI) is used to diagnose the breast tumors. In general, the morphology can reflect the malignancy or benignancy of a tumor. Several conventional shape features are used to extract the shape information. Furthermore, a fitting ellipsoid is built by the tumor contour and then some morphological features are proposed by comparing the tumor contour with its corresponding ellipsoid. Besides, the 3-D texture information based on the grey level co-occurrence matrix is also considered a diagnosis feature and also be used in this paper for a comparison. In our experiments, 95 pathology-proven cases, which contain 44 benign tumors and 51 malignant ones, are used to test the accuracy of several different feature categories and our proposed computer-aided system. From the experimental results, we could find that the morphological features have better performance than the texture features. Moreover, the proposed 3-D morphological features could achieve a high performance with the accuracy, sensitivity, and specificity being 88.42%, 88.24%, and 88.64%, respectively.