Hierarchical Masses Detection Algorithms Based On SVM In Mammograms

Breast cancer is one of the most common malignant diseases, and the early diagnosis and treatment can significantly increase the chance of survival for patients. Mammography is the first choice to diagnose breast cancer. However, features of early breast cancer in mammograms are unconspicuous, so even the seasoned doctor cannot easily discover all of the possible disease in time. With the rapid progress of physic and computer technology, detection of masses and calcifications on mammograms has been a hot research field in early detection of breast cancer. A good computer aided detection (CAD) system, which can effectively avoid error and miss in diagnosis due to the eyestrain or negligence of human, could help doctors recognize micro disease better in medical images.Masses are the major indications of breast cancer on mammograms. For this purpose, this paper presents a series of new methods to detect masses automatically. Based on the morphological analysis, a new image enhancement method is proposed, which can effectively suppress the background and enhance the features of masses on mammograms simultaneously. Then the plump seed regions can be extracted from the images using features such as gray or contrast in the enhanced and original images. A novel method inspired by vague set, is introduced to extend the fuzzy region growing to a vague version, which gauranttees the completeness as well as stability of region growing. Since some masses have low density, the detection on one gray scale cannot reach them. So a hierarchical detected method is developed, which can detect the unconspicuous masses effectively. Owing to a great deal of false positives lie in ROI, the SVM classifier is designed to distinguish masses from nomal areas with good detected result. To improve the ture positive while reduce false positive, the relevance feedback is introduced to filter out the number of false positives.The experimental results show that the proposed detection algorithms can obtain good detection result. It is believed that these detection algorithms will be of extensive application prospect.