الفهرس 
IntroductionOnly 14 pages are availabe for public view
 |  | from | 126 |  |  |
| | | from | 126 | | |
Abstract
Circumscribed masses in mammograms are an important early sign of breast cancer. This thesis presents a computeraided diagnosis (CAD) system for the automatic detection of circumscribed masses in digitized mammograms. The proposed system consists of three main steps; first, images were preprocessed to segment the breast tissue from the background of the image and to enhance the image contrast. This was achieved using histogram equalization and thresholding for segmentation and a wavelet enhancement technique. Second, two separate sets of features are extracted from each mammogram; four statistical features and seven texture features derived from the cooccurrence matrix. Third, the discriminatory power of these features is analyzed using a neural network and a fuzzy neural network. The method is applied to a database of 22 mammograms (MiniMIAS database) containing 24 masses. Results show that the proposed system gives quite satisfactory detection performance. In particular 100% true positive detection rate is achieved at the cost of 30% false positive when statistical features and neural networks are used, and 100% true positive detection rate is achieved at the cost of 20% false positive when texture feature derived from cooccurrence matrix at angle zero and a neurofuzzy classifier are used. These results indicate that the selected techniques of image processing, feature extraction and artificial intelligence are a promising approach for improving the accuracy of classification in CAD applications.