目前X光攝影是最有效偵測乳房腫瘤的方法。本論文使用的實驗影像資料,是由歐洲乳房X光影像分析學會(The Mammographic Image Analysis Society) 所提供的 MIAS Mini-Mammographic Database。在這組資料庫中,有161位受試者的左右乳房影像,共322張數位影像。本研究使用其中的117張有腫瘤的影像做為實驗用影像。 本研究總共選用了126個紋理特徵,分別為灰階強度統計4個、空間灰階相關特徵44個、紋路頻譜9個、紋路特徵編碼48個、區域性灰階相依紋理特徵5個、灰階長度紋理特徵16個。在常用的紋理特徵之外,本研究特別加入了小波特徵及近年來較新興的曲波特徵一起分析。另外為了找出最具代表性的特徵,使用了費雪線性區分法,對所有特徵做排名,可以挑出重要特徵並把特徵維度降到10~15個。分類器系統則使用支持向量機。 在分類階段,使用未經特徵選取於所有類型組織下的126個一般紋理特徵,其分辨率為98.4%。使用特徵選取可以保留有效特徵並將特徵大幅降低至10個,其分辨率仍可達92.9%。當加入小波特徵後可將分辨率提升至98.2%,再加入曲波徵後更上升至99.1%。 在偵測階段,本研究討論不同尺寸的遮罩分別為32x32、64x64及128x128,其中64x64的效果最好。絕大部分的腫瘤可以被找到,尤其以脂肪型組織最為明顯,偽陰率(False negative)為0.12/image;緻密型及脂肪腺體型則較不容易偵測到,其偽陰率分別為0.17/image及0.16/image。 本系統可輔助醫師在臨床上的診斷,藉由提高診斷正確率,使病人降低生命的危害。
Nowadays, mammogram is the most effective breast tumor detection method. In this research, we used the Mini-Mammographic Database provided by the Mammographic Image Analysis Society (MIAS). There are 322 digital images of left and right breasts from 161 patients. We used 117 of them that contained tumor image for experiments. This research employed totally 126 texture features, including 4 from Gray Level Histogram, 44 from Spatial Gray Level Dependence, 9 from Texture Spectrum, 48 from Texture Feature Coding Method, 5 from Neighboring Gray Level Dependence Textural Feature, and 16 from Gray Level Run-Length Textural Features. Apart from the commonly used texture features, we recruited wavelet features and the curvelet features which were developed in recent years. In the study, to select the most representative features, we used the Fisher’s linear discriminant analysis to rank the features and selected the most discriminative 10~15 features. The support vector machine was employed as the classifiers. In the classification apart, using the entire texture features, an accuracy is of 99.4% was attained in recognizing tumors in different types of tissues. By using the feature selection method, we can extensively decrease the amount of features and keep the effective features. An accuracy of 92.9% was retained by using only 10 effective features. Moreover, adding wavelet features enhanced the accuracy up to 98.2%. Further adding curvelet features elevated the accuracy to 99.1%. In the detection part, we first studied the influence of matrix sizes. Among the three matrix sizes, 32x32, 64x64 and 128x128, matrix of size 64x64 contributed to the best results. Most of the tumors could be detected, especially tumors embedded in the fatty tissues. The false negative was 1.31/image. Comparatively, tumors were relatively difficult to be detected in dense glandular tissue. The false negative for dense glandular and fatty glandular tissue were 1.67/image and 1.38/image, respectively. The proposed method was demonstrated to be effective in assisting doctors in diagnosizing breast tumors. By enhancing the diagnosis accuracy, the mortality rate of the patients can be reduced.