本文為馬來西亞馬來亞大學（作者：JAWAD NAGI）的電腦科學碩士論文，共355頁。

乳腺癌是最常見的癌症之一，也是導致女性死亡的主要原因。乳腺攝影是目前最有效的乳腺癌影像學檢查方法，可用於鑑別異常癌細胞。研究顯示，在當前的乳腺癌篩查中，大約15%到30%的乳腺癌病例被放射科醫生漏診。隨著數字影象處理技術的進步，放射科醫師將有機會減少這種誤診的概率，從而提高確診率。

數字乳腺X線攝影已成為乳腺癌最有效的檢測手段。本研究的目的是通過減少誤分類的癌症數目，提高影象處理和機器學習技術的診斷準確性，以優化數字乳腺攝影中惡性和良性腫瘤的判斷。在本研究中，數字乳房攝影影象是從2008年至2010年在馬來亞大學醫學中心（UMMC）接受治療的馬來西亞患者獲得的；該資料庫由密集、脂肪和脂肪腺乳房構成的標準影象組成，根據檢查結果分為正常、良性和惡性三類。

本研究應用影象處理技術對乳腺X光影象進行增強，以實現乳癌的計算機檢測。用於乳腺X光影象處理的演算法包括形態學運算和閾值處理技術。由於數字乳腺照片中的胸肌會影響檢測結果，因此應該在乳房X光照片中進行抑制。本研究採用種子區域生長技術，從胸肌中分離出乳房組織。

使用真實(GT)資料和放射學家對乳腺攝影資料集解釋得到的標記，選擇分割後圖像的惡性和良性樣本，並標記為感興趣區域(ROI)或異常區域(樣本)。利用灰度共生矩陣（GLCMs）從ROI樣本中提取紋理特徵；為了對惡性和良性樣本進行模式分類，利用支援向量機（SVM）對紋理特徵的最優子集進行建模。SVM使用三分之二的總樣本進行訓練，其中剩餘的三分之一樣本用於測試和驗證。本文利用接收機工作特性（ROC）分析，結合靈敏度、特異性和曲線下面積（AUC）等效能指標，測量了所開發系統的二元分類精度。為了進行對比研究，本文評估了SVM以外的機器學習演算法，即人工神經網路（ANN）。本系統所獲得的實驗結果證明了對乳腺癌的自動檢測是有益的，該技術可以提高放射科醫師在乳腺癌診斷中的診斷準確性和一致性。由此研製的乳腺癌自動檢測系統將作為放射科醫師手動檢測後的第二判圖者，我們相信這將有助於提高放射科醫師的確診概率。

Breast cancer is one of the most common kinds of cancer, as well as theleading cause of mortality among women. Mammography is currently the mosteffective imaging modality for the detection of breast cancer and the diagnosisof the anomalies which can identify cancerous cells. Retrospective studies showthat, in current breast cancer screenings approximately 15 to 30 percent ofbreast cancer cases are missed by radiologists. With the advances in digitalimage processing techniques, it is envisaged that radiologists will haveopportunities to decrease this margin of error and hence, improve theirdiagnosis. Digital mammograms have become the most effective techniques for thedetection of breast cancer. The goal of this research is to increase thediagnostic accuracy of image processing and machine learning techniques foroptimum classification between malignant and benign abnormalities in digitalmammograms by reducing the number of misclassified cancers. In this research,digital mammography images are obtained from Malaysian patients who are treatedat the University of Malaya Medical Centre (UMMC) from 2008 to 2010. Thisdatabase consists of standard images of dense, fatty and fatty-glandularbreasts, which are classified into three categories: normal, benign andmalignant, using the results obtained from biopsies. Image processingtechniques are applied in this research to enhance the mammogram images for thecomputerized detection of breast cancer. Image processing algorithms used formammogram image processing include morphological operations and thresholdingtechniques. As the pectoral muscle in digital mammograms can bias the detectionresults, it should be suppressed from the mammograms. This research employs aseeded region growing technique for the segmenting the breast tissue from thepectoral muscle. Malignant and benign abnormalities are selected from thesegmented images using the Ground Truth (GT) data and markings obtained fromthe radiologists’ interpretation of the mammography datasets, which correspondto the Regions of Interest (ROIs) or abnormal regions (samples). Texture basedfeatures are extracted from the ROI samples using Gray Level Co-OccurrenceMatrices (GLCMs). For the purpose of pattern classification between malignantand benign samples, the optimum subset of texture features are modeled using aSupport Vector Machine (SVM). The SVM is trained using two-thirds of the totalsamples where the remaining one-third of samples are used for testing andvalidation. The binary classification accuracy of the developed system ismeasured using the Receiver Operating Characteristic (ROC) analysis withperformance measures such as sensitivity, specificity and the Area Under theCurve (AUC). To perform a comparative study, machine learning algorithms otherthan the SVM, namely, Artificial Neural Networks (ANNs) are evaluated in thisresearch. The experimental results obtained from the system developed in thisresearch prove to be beneficial for the automated detection of breast cancer.The proposed technique will improve the diagnostic accuracy and consistency ofthe radiologists’ image interpretation in the diagnosis of breast cancer. Theresulting computerized breast cancer detection system will subsequently act asa second reader after the manual detection by the radiologist and it isbelieved that this would aid the radiologist in the mammogram screeningprocess.

1 引言

2 數字化乳腺攝影

3 計算機輔助檢測的要素

4 模式識別與特徵選擇

5 建模框架

6 實驗結果與討論

7 結論與未來研究方向

附錄A 資料建模與分析

附錄B SVM訓練與驗證

附錄C LIBSVM版權提醒

附錄D 發表論文列表

下載英文原文地址：

更多精彩文章請關注微訊號：