請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/58929
標題: | 超音波電腦輔助腫瘤偵測與可靠性診斷 Computer-aided Lesion Detection and Robust Diagnosis for Breast Ultrasound |
作者: | Min-Chun Yang 楊閔淳 |
指導教授: | 張瑞峰 |
關鍵字: | 乳癌,可攜式超音波,簡易貝式分類器,腫瘤偵測,可靠式電腦輔助診斷, Breast cancer,portable ultrasound,naive Bayes classifier,lesion detection,robust computer-aided diagnosis, |
出版年 : | 2013 |
學位: | 博士 |
摘要: | 根據美國癌症協會的最新研究報告,乳癌是女性癌症死亡的第二大主因。因乳癌的切確形成原因還未能完整的了解,但我們能夠從早期的乳癌偵測與診斷有效用並有效率的增加臨床治療成功率同時降低不必要的腫瘤穿刺。超音波成影技術是目前其中一種廣泛使用來偵測與分類乳房腫瘤的非侵入式治療法。目前,有許多的研究學者致力於開發電腦輔助系統(如 CADx 和 CADe)來協助醫生解釋超音波影像如腫瘤偵測與診斷等的輔助工具。相對於使用傳統笨重且龐大的超音波掃描機器來做超音波影像的篩檢,可攜式的超音波設備更能提高病人的檢查比率。在本研究中,我們使用可攜式超音波機器Terason t3000 (Terason Ultrasound, Burlington, MA, USA) 並使用手持探頭擷取來源超音波影像以做進一步的分析。為了有效率的處理巨量資料的議題以利發展實用的全乳房腫瘤偵測臨床應用軟體,我們提出了快速簡易貝氏分類法來處理像素的分類問題與二階段的腫瘤篩選機制來挑出真正可疑的腫瘤同時壓低系統的偽陽性。我們所提出的系統能達到 93.94% 的偵測準確率及每百張超音波影像4.22個偽陽性誤判腫瘤,此系統能夠很有效的幫助醫生重新檢視系統所偵測出的少許可疑的腫瘤以利快速的完成第二次乳房檢查的進行。同時為了減少腫瘤的穿刺,許多的研究人員亦致力於使用以GLCM為基礎的紋理特徵開發電腦輔助診斷以協助腫瘤的診斷。但是,這些紋理特徵的研究分析並沒有考量到超音波機器上的參數設定會影響資料的取樣而造成跨機器平台診斷效能的變異。因此,我們的研究專注於使用超音波影像不受到亮度影響的轉換方法以開發一個可靠的電腦輔助診斷的系統。當我們從轉換後的超音波影像中擷取多解析度的特徵後,診斷的效能在此不受到亮度影像的特徵下比目前在發展中的紋理分析的結果都還要更好。我們做了一些實驗並進行接收者操作特徵分析 (ROC),我們得到三台機器診斷的效能在接收者操作特徵分析下的面積(AUC)分別是0.918 (95% 信賴區間為 0.848 到 0.961), 0.943 (95% 信賴區間為 0.906 到 0.968) 以及 0.934 (95% 信賴區間為 0.883 到 0.961)。實驗的結果顯示使用數值排序法來進行紋理分析較不易受到跨機器平台的影響並且適合用來設計與開發一個可靠的臨床腫瘤診斷的應用軟體。 Breast cancer is the second leading cause of death for female as latest reported from the American Cancer Society (ACS). Since the exact causes of the disease remain unknown, early detection and diagnosis can effectively and efficiently increase successful clinical treatment while reducing unnecessary biopsies of breast masses. Ultrasound imaging is one of widely used non-invasive modality to detect and classify abnormalities of breast lesions. Recently, many researchers are dedicated to develop computer-aided systems (i.e., CADx and CADe) for breast cancer detection and diagnosis to aid radiologists in interpreting breast ultrasound (BUS) images. Compared to traditional heavy and large ultrasonic machines for screening BUS images, portable PC-based breast ultrasound (BUS) imaging systems can be adopted to improve the patient throughput. Herein, we use PC-based ultrasound machine Terason t3000 (Terason Ultrasound, Burlington, MA, USA) with free-hand probe to acquire the source BUS images for later analysis. To effectively address the big data issue while developing clinical applications for whole breast lesion detection, the proposed pixel classification based on naïve Bayes classifier is used to categorize the normal or lesion objects and two-phase lesion selection scheme is adopted to pick out all the suspected lesions with a lower false-positive rate. The proposed system present 93.94% with 4.22 false-positive per hundred slices and is effective for the radiologists to perform the second examination by merely reviewing the detected lesion objects of the proposed system. In order to reduce the biopsies of the breast masses, many researches devote to develop computer-aided diagnostic system using GLCM-based textural features for tumor diagnosis. Nonetheless, these texture analyses did not consider varied parameter setting of different ultrasonic devices might result in large variation of diagnostic performances across the sonographic platforms. Therefore, we aim at developing a robust computer-aided diagnostic system for BUS images based on the invariant gray-scale transform (i.e., ranklet transform). While multi-resolution features are extracted from the ranklet transformed BUS images for texture analysis, the diagnostic performances with the invariant texture features outperform those with state-of-art texture analyses. We conducted experiments in terms of receiver operating characteristic (ROC) analysis, the AUC values derived from the area under the curve for the three databases are 0.918 (95% confidence interval [CI], 0.848 to 0.961), 0.943 (95% CI, 0.906 to 0.968) and 0.934 (95% CI, 0.883 to 0.961), respectively. The experiments reveal the texture analyses using ranklet transform are less sensitive to different ultrasonic devices and properly adopted for designing a robust system for tumor diagnosis. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/58929 |
全文授權: | 有償授權 |
顯示於系所單位: | 資訊工程學系 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-102-1.pdf 目前未授權公開取用 | 2.54 MB | Adobe PDF |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。