Breast cancer is one of the leading causes of deaths among women. Mammography is currently the best method for early detection. Using low dose x-ray in mammography, the breast tissue type and different kinds of lesions make the detection of lesions in mammograms very ambiguous and tedious work. Early detection is the most effective way to reduce the mortality rate. Our main aim in this thesis is detection and recognition of tumors in mammograms. Mammograms usually have large size so the processing of the entire mammogram takes a lot of time. To reduce the size and therefore the processing time and also decreasing FPR, a two-step algorithm is used. At the first step some unimportant regions such as background and pectoral muscle are eliminated and at the second step an ROI detection algorithm is proposed which extracts the most likely regions to tumors. To recognize the tumors in detected regions, some features are extracted from each region. To find the most effective features for tumor detection, several data mining, features extraction and feature selection methods are used and compared. At the end a very effective method is proposed by mixing the data extracted from co-occurrence matrix and PCA. To increase the performance and reduce the number of features a GAs based algorithm is proposed. SVM is used as our final classifier, because it has the best results in comparison with other tools in our application. Finally an approach is suggested to recognize if the tumor is benign or malignant which is based on finding the border of tumor, opening the border around its center of gravity and extracting some features such as fractal dimension. The reached area under ROC curve using proposed co-occurrence features, PCA and feature selection based on GAs is 0.97. The TPR using SVM is 97.3% and FPR is 16.65%. Experimental results show that the performance of proposed methods is better than other methods.
