Diagnosis of breast cancer with image processing techniques

Breast cancer is the most frequent type of cancer that causes death in women both in the world and in our country. When we look at cancer statistics in our country, breast cancer constitutes about 25.6 % of all female cancers. Breast cancer is seen in one out of every 8 women in Turkey. The most effective method for reducing deaths due to breast cancer is early diagnosis and treatment. The most common method used for early diagnosis is mammography. Mammography is a special X-ray technique designed for breast cancer. Images obtained with mammography can be used to determine even small changes that cannot be noticed by manual examination. Interpretation of mammography is done by expert radiologists. But fatigue, workload, lack of human vision, etc. may cause radiologists to misinterpret mammograms. Nowadays, automatic detection systems are being developed to help radiologists in order to reduce the effects that may lead to incorrect or incomplete interpretation in the diagnosis of breast cancer. These systems are called Computer Aided Diagnosis (CAD) system. CAD software uses a variety of image processing algorithms to detect abnormalities in the breast, and as a second opinion to radiology specialists in the diagnosis and diagnosis of breast cancer. The final decision is made by the radiologist. In this study, a CAD system was developed to detect and classify abnormalities in mammograms. This system mainly consists of five sections. These are respectively; pre-processing, segmentation, feature extraction, feature selection and classification. In the pre-processing phase, are used median filter, biorthogonal wavelet analysis, anisotropic diffusion method, adaptive histogram equalization method for image enhancement and noise reduction. In addition, a new method developed within the scope of this thesis has been used to remove the pectoral muscle. In order to detect the suspicious areas, w-BSAFCM which is an image clustering algorithm developed within the scope of this thesis, Otsu N threshold, Havrda & Charvat entropy methods are used together. Gray level co-occurrence matrix, wavelet transformation and curvelet transformation methods are used for feature extraction. Linear Discriminant Analysis method is used in the feature reduction phase. In the classification phase, artificial neural networks, support vector machine (SVM) and K- nearest neighborhood classifiers are used. In addition to the MIAS database frequently used in the literature, the proposed method and developed algorithms for the thesis study were also tested on images of the INBREAST database and the results obtained were compared with other studies in the literature. In experimental studies, mammogram images were first classified as normal and abnormal by the developed method and then abnormal mammograms were classified as benign and malignant among themselves. A total of 246 mammograms were used for the thesis work. The best results were obtained with the proposed methods and the curvelet transform for feature extraction by using 70 normal and 70 abnormal mammograms selected from the MIAS database, and 53 normal and 53 abnormal mammograms selected from the INBREAST database. The dimensionality was reduced by linear discriminant analysis of the extracted features by the curvelet transform and the classification performance was obtained as 100 % when the linear analysis and curvelet transform were used together. In addition, gray level co-occurrence matrices and wavelet analysis results for the classification of mammograms are presented comparatively.