請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/59975
標題: | 使用材質資訊於全乳房自動超音波的腫瘤偵測 Tumor Detection Based on Texture Information in Automated Breast Ultrasound |
作者: | Shih-Chun Lin 林士淳 |
指導教授: | 張瑞峰(Ruey-Feng Chang) |
關鍵字: | 全乳房自動超音波,乳癌,電腦輔助乳房腫瘤偵測,偽陽性篩除,多元邏輯斯迴歸,組織分類,分水嶺切割, automated breast ultrasound,breast cancer,computer-aided detection,false positive reduction,multinomial logistic regression,tissue classification,watershed segmentation, |
出版年 : | 2017 |
學位: | 碩士 |
摘要: | 全乳房自動超音波是一種新的乳房腫瘤偵測儀器,與傳統手持式超音波相比, 全乳房自動超音波具有幾項優點,包括更高的再現性和更少的操作者依賴性。然 而,一組全乳房自動超音波影像包含數百張二維影像,若逐張檢查,醫師需要花費 很多時間,因此電腦輔助乳房腫瘤偵測系統被設計來解決檢查全乳房自動超音波 影像的耗時問題。先前一個基於分水嶺切割的電腦輔助乳房腫瘤偵測系統能夠達 到腫瘤偵測率 100%,其伴隨著一些非腫瘤區域被誤判為腫瘤。由於大部分的腫瘤 位於乳腺區域的內部和附近,乳房的解剖學資訊可以被使用於電腦輔助乳房腫瘤 偵測系統的偽陽性篩除上。因此,本論文提出一個基於解剖學的偽陽性篩除模型以 提升先前電腦輔助乳房腫瘤偵測系統的效能,利用每一個區域中量化的空間和材 質特徵設計一個邏輯迴歸模型來從全乳房自動超音波影像取出腺體區域,利用電 腦輔助乳房腫瘤偵測系統所偵測出來的腫瘤區域與腺體區域作疊合,再根據疊合 率使非腫瘤區域被誤判成腫瘤的個數有所下降。實驗中所用到的測試資料包含 43 個良性以及 39 個惡性腫瘤。根據結果,本系統在腫瘤偵測率為 100% (82/82), 90% (74/82), 80% (66/82)的情況下,平均每一個全乳房自動超音波掃描會有 3.15, 1.90, 1.42 個非腫瘤區域被誤判成腫瘤,與先前的電腦輔助乳房腫瘤偵測系統相比 非腫瘤區域被誤判成腫瘤的情形下降了 67.49%。總結,本論文所提出之基於解剖 學的偽陽性篩除模型,在針對全乳房自動超音波掃描作腫瘤偵測時,能夠在不影響 腫瘤偵測率下有效地降低非腫瘤區域被誤判成腫瘤的情形。 Automated breast ultrasound (ABUS) is a new screening modality. Compared to the conventional handheld ultrasound, ABUS has several advantages, including higher reproducibility and less operator dependence. However, an ABUS image volume includes hundreds of 2-D slices, and the physicians need a large amount of time to review the volume image. Hence, the computer-aided detection (CADe) systems were proposed to solve the long-time review issue. A CADe system based on watershed transform can achieve the sensitivity of 100% with several false positives (FP) in the previous study. Because most tumors are located inside or close to the mammary gland regions, the information of breast anatomy can be used for reducing the FPs of CADe. Therefore, this study proposed an anatomy-based false positive reduction (FPR) method to decline the FPs/pass with the same sensitivity as the previous CADe system. A logistic regression model using quantitative spatial and texture features was designed to extract glandular regions from ABUS images. The FPs would be reduced according to the overlapping ratio between glandular regions and suspicious regions detected by CADe system. The collected database comprised 43 benign and 39 malignant tumors. As a result, the proposed anatomy-based FPR achieved sensitivities of 100% (82/82), 90% (74/82), 80% (66/82) with FPs/pass of 3.15, 1.90, and 1.42, respectively. A decrease of 67.49% in FPs/pass was achieved in comparison to the previous CADe system. In conclusion, the proposed anatomy-based FPR could reduce FPs with the same tumor detection rate in ABUS images. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/59975 |
DOI: | 10.6342/NTU201700132 |
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-106-1.pdf 目前未授權公開取用 | 2.29 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。