中文摘要

在眾多類型的乳癌診斷程序中，醫學超音波乳癌影像診斷為接受度最高的及時診斷工具，因此超音波掃描儀成為臨床上醫生早期診斷乳癌的利器。然而，由於超音波腫瘤影像含有大量的斑點、 雜訊及組織紋理，容易因為不同廠牌型號的超音波掃描儀輸出的影像相異，或因操作的醫師經驗差異產生不同診斷結果的情形，進而導致醫療資源的濫用。因此近年來，本計畫主持人與多位學者及醫師合作完成數個乳房超音波腫瘤電腦輔助診斷系統，得到了卓越的診斷結果，但是此系統的輸入資料需要醫師手動擷取的腫瘤區域，為了降低系統的應用門檻，協助臨床醫生快速描繪腫瘤的正確輪廓，於是結合紋路分析技術、類神經網路模式、分水嶺轉換自動描繪乳房腫瘤的輪廓，使系統可自動描繪出近似於豐富經驗的醫師所描繪之腫瘤輪廓。由於上述兩個主題深具實用性，可造福更多婦女同胞，所以在本計畫中，將針對乳癌超音波電腦輔助診斷系統作一整合性研究，以期完成自動電腦輔助診斷系統。
在本計畫中，我們將分成三個階段進行，首先針對紋路特徵使用主成份分析降低特徵向量的維度，並採用影像搜尋技術將乳房腫瘤分類，做為判斷腫瘤良惡性之依據；第二階段將使用數位影像加強技術減低不同廠牌型號的超音波掃儀之輸出差異，加強腫瘤影像紋理特徵診斷的效率及實用性；第三階段為簡化分水嶺轉換輪廓描繪法之複雜度，以超音波影像紋路為主軸，探討乳房超音波影像內的組織差異，利用不同組織的影像特性來合併腫瘤內部的過度切割區域，以期加強腫瘤輪廓描繪的速度及一致性，並提供自動擷取的腫瘤區做為乳房腫瘤電腦輔助診斷系統的輸入。最後結合前述所發展之技術，建構乳房超音波腫瘤輪廓描繪、診斷的自動化系統。

關鍵詞：乳房超音波、紋路分析、分水嶺轉換、乳癌診斷、腫瘤輪廓描繪、腫瘤偵測

Abstract

Medical ultrasound is a certified convenient tool for detecting and diagnosing breast tumors, particularly palpable tumors. Due to the ultrasonic examination would not cause any side effect upon patients’ body. The cost-effectiveness and portability of this facility are particularly important in smaller hospitals, in which the equipment is useful in conducting complex medical imaging in a timely manner. The ultrasonic image is also an efficient instrument of the clinical physicians to diagnose the breast tumors at an earlier stage. However, speckles, noise, and tissue-related textures always consist in sonographic images, physicians without clinical experience may make a miscarriage of diagnoses. Hence, we have proposed a number of computer aided diagnosis (CAD) systems for breast ultrasound and given remarkable results. In these systems, the region of interesting (ROI) in the ultrasonic image is manual sketched by the experienced physicians. However, the automatic contouring may assist physicians without experience in making a correct diagnosis. Textural analysis, neural network, and watershed transformation techniques were employed to automatically identify the contour of breast tumors.
In this project, we will design a complete CAD system by using the previous studies. There are three stages to achieve the aim. First, the principal component analysis (PCA) will be utilized to diminish the dimension of textural feature vector. We will utilize the image retrieval methods to diagnose breast tumors by using the automatic extracted ultrasonic sub-image of the ROI. Second, image enhancement techniques will used to make no difference between sonographic images that acquire form diverse ultrasound equipments. Third, the watershed transform will be performed to identify tumor’s contour in the ultrasound textural image. A texture-based region merging method absorbs the regions in a tumor according to the textural similarity function. We will try to identify the similar contour with the physician for the breast ultrasonic images more effectively. Finally, the techniques of the three stages will be combined for achieving an automatic CAD for 2-D breast ultrasound images.

Keywords: Breast ultrasound, Texture analysis, Breast cancer diagnosis, Tumor contour approximation, image segmentation, computer aided diagnosis