Research On Detection Methods Of Breast Mass In Mammography

Breast cancer is one of the most common malignancies with high incidence and mortality rates among older women. To increase the cure rate and reduce the mortality from this disease, early detection and treatement of breast cancer is the important premise. Mammography is considered as the most reliable and cost-effective method for early screening and diagnosis of breast cancer. But the problems, such as fatigue, time-consuming, higher rates of misdiagnosis and missed diagnosis, will occur if the detection of breast cancer is solely performed by a single radiologist. Computer-aided detection and diagnosis system, which is based on computer image processing and artificial intelligence techniques, has the potential to improve the rate of early detection, and make diagnosis results more objective and accurate.This dissertation focuses on computer-aided mass detection in a single mammography, which aims to improve the performance of mass detection. The main contents are as follows:breast region segmentation, suspicious lesion location and segmentation, feature extraction and selection, classification of suspicious mass, and validation of effectiveness of the proposed algorithm.For breast region segmentation, the tendency of Otsu threshold is analyzed theoretically and the conclusion that Otsu threshold will split the object with a larger variance is obtained. Based on this conclusion and by considering the characteristic that the variance of breast region is larger than that of background region, we propose an improved Otsu method with constraining the gray level range to segment the image. After that, the breast region is extracted by morphological operation on the segmentation result. Experiments show that the segmentation results are excellent.To locate suspicious lesions, we first compare systematically the match performance of a number of templates with different sizes and morphologies. Based on the characteristics that the gray value of a mass center is relatively large, the pixel values reduce slowly from the center to outer, the mass shape nearly like a circle, the similarity between a mass and a template is quite large when their sizes are approximately same, and the smaller template has a higher sensitivity while the larger template has a lower false positive rate, we design a multiple scales template matching method to improve sensitivity and reduce false positive rate at the same time.To segment suspicious mass, we propose a dynamic programming algorithm with constrained search range. At first, we apply plane fitting technology based on the least-squares method to correct the gray-level distribution of ROI (region of interest), then calculate the cost which is constructed by the radial gradient, the deviation of pixel value from expected gray level (intensity), and the distances between the nodes. The weight values of the three components are optimized by a particle swarm optimization algorithm (PSO). By using the strategy of constraining search space, we address the problem that the path with the minimum cost in ROI is not the best mass boundary. Experimental results show the validity of our method by comparing our segmentation results with gold standards in terms of the overlap percentage and Hausdorff distance.On feature extraction and selection, we extract gray features, texture features and morphological features, and then use a stepwise discriminant analysis method to select features. On classifier, we develop Fisher classifier and Logistic classifier. At last, we use leave-one-out test method to evaluate the performance of the CAD system, and draw the corresponding ROC and FROC curves.The experiments on the dataset seleted from DDSM show that our computer-aided mass detection system has a high performance and lays a good foundation for the use of the system in clinical applications.