| This thesis compares different methods that could be used to construct a computer aided diagnosis (CAD) system that analyzes mammograms. Such systems have many steps, but this thesis focuses on feature extraction, feature selection, and classification. The main comparison is between the simplified Rubber Band Straightening Transform (SRBST) and the Onion Transform, which are used to extract texture features. Another comparison is between 2D and 3D co-occurrence matrices. Next, features are selected using a greedy algorithm. This section mainly compares the effectiveness of Receiver Operating Characteristic (ROC) and Class Overlap Rating (COR). Also evaluated are the effectiveness of Linear Discriminate Analysis (LDA) and the sort order of features. Then the selected features are used to classify the lesions. In this part, Nearest Mean, Nearest Neighbor, and Maximum Likelihood are compared. The results are then used to determine the best combination of methods to use in a CAD system. |
