- 學位論文
數位乳房攝影電腦輔助偵測系統之整合
The Integration of Computer-Aided Detection System of Digital Mammography Image
摘要
乳癌為國人十大死因之一，腫塊、微鈣化及組織牽扯等病徵皆為乳癌發生初期的重要病徵，但以肉眼觀察影像，較不易發現。因此本研究使用影像處理的技術整合一套數位乳房攝影影像之乳癌電腦輔助偵測系統，用來輔助特徵點的可辨識度，降低醫師在診斷時的誤判率。 本研究所整合之系統提供腫塊位置偵測、微鈣化位置偵測及組織牽扯偵測和分類等功能。偵測腫塊位置方面，利用向量域收斂濾波器與區域成長法找出腫塊位置；而偵測微鈣化位置方面，對鈣化影像進行小波轉換並使用動態區域選取的方式進行組織分群；而偵測組織牽扯方面，也是利用收斂向量濾波以及去除胸大肌結合平均閥值能有效的排除影像中之誤判區域數量。系統開發時，先以假體影像確認系統之運算的正確性。而後再使用經確診之64張的乳房FFDM影像分組來進行系統之訓練與驗證。再擷取訓練後之特徵與紋理參數進行統計分析，將分析結果先利用獨立T-test進行篩選後，找出具有鑑別能力的特徵參數並進行良惡性支持向量機(Support Vector Machines)與Leave-one-case-out的分類，並與醫師判讀進行比對。最後在研究中也開發出一套友善的程式介面供使用者操作。 在假體測試時，可發現其在對於尋找腫塊中心、增強微鈣化影像以及擷取乳房組織的位置上均可達良好的測試效果，且能符合系統之需求。於臨床影像腫塊判讀中，在每張影像上皆可偵測到正確腫塊位置，依照病例偵測之靈敏度為100%，系統經由SVM良惡性分類之訓練組與測試組的效能其Sensitivity，Specificity，Accuracy均達100%，即Kappa值為1，而每張影像的系統平均誤判數量為1.92個；在微鈣化偵測上，系統最佳分割閥值0.3% 時，依照病例偵測之偵測靈敏度為100%，由良惡性分類之訓練組與測試組結果顯示，其Sensitivity，Specificity，Accuracy均達100%，即Kappa值為1，每張影像的系統平均誤判數量為1.92個；而在組織牽扯偵測上，在每張影像上皆可偵測到正確組織牽扯位置，依照病例偵測之偵測靈敏度為100%。良惡性之分類訓練組與測試組結果顯示，其效能為Sensitivity=1，Specificity=0.857，Accuracy=0.818，Kappa=0.61，但每張影像的系統平均誤判數量為11.08個，其結果能符合現階段臨床之需求。 本研究已整合此數位乳房電腦輔助診斷系統可以分別針對乳房腫塊、微鈣化及組織牽扯進行偵測及分類，不僅可以幫助醫師對患者做進一步的病理判斷，並提供客觀的資訊給經驗較不足的醫師進行輔助診斷。
並列摘要
Breast cancer is one of the top ten causes of death, the mass, micro-calcification and architectural distortion (AD) are important early symptoms of breast cancer occurred, but it can be ignored easily. The aim of this project is to integration of CAD system for digital mammography for breast detection. The system provided the mass location detection, the micro-calcification location detection and the architectural distortion location detection and provided classification function. The Otsu’s method was used to find the best threshold to remove the background and noise of image by region growing method during preprocessing stage. For the detection of masses and architectural distortions, Convergence Index Filter method was used to enhancement breast tissue image and to locate the mass. For the detection of micro-calcification, enhanced the image contrast by using wavelet transform and grouped is performed using dynamic region selection to locate of suspicious micro-calcifications. For the detection of architectural distortion, we used manually method to segment pectoral muscle and average threshold can effectively decrease the number of ROIs of the image. First, we used prosthesis image to prove the system is correct. Next we used 64 FFDM image to group for training and validation of the system. Then we extracted the training feature and texture parameters for statistical analysis. Features of image are selected through t-test statistical analysis. After evaluating the results by the independent t-test, the effective features were selected and severed as inputs in the support vector machines (SVM) and Leave-one-case-out. We compared with the doctors interpretation. Moreover, a user friendly interface also developed in this study. In the prosthesis test, system could search for the center of the mass, enhanced micro-calcifications image and captured breast tissue, system had a good test results, and conform to the system. The results show that the sensitivity of this detection rate is nearly 100% as if a patient has both MLO and GC view images. For both the training and testing data of mass and micro-calcification, by SVM classification with 100% for sensitivity, specificity, accuracy, and kappa=1; and for the training and testing data of architectural distortion of system, by SVM classification with sensitivity=1, specificity=0.857, accuracy=0.818, and kappa=0.61. But the average number of false positives of mass, micro-calcification and architectural distortion for system is 1.92, 1.92 and 11.08, respectively. These results can meet the requirement of physician. The developed computer-aided diagnosis system is not only help physicians great advance in patients diagnosing breast cancer, but also assist a junior physician through more objective information to make a diagnosis.