Detection of masses in X-ray mammograms
A circular filter having two disjoint regions, an inner disk and outer ring, is applied in two different nonparametric ways for identifying abnormal masses in digitized X-ray mammograms. Data points within an annulus are used to estimate a parameter of interest. In the first method, the number of texture features is extended from one to three, including standard deviation. The two additional features are skew and kurtosis. They are mapped to the center of a fixed size template and used to measure local texture characteristics within the mammograms. Feature images are created using histogram equalization and gray level scaling. The mapping results in a new image containing a desired “halo” type response that identifies regions with changing texture. For the five types of abnormalities combined, the desired response was obtained in 87% of the images using standard deviation, 84% by skew and 80% by kurtosis. A second visual test, using twenty-two subjects, was used to evaluate the performance of the circular template. These results show that standard deviation, skew and kurtosis visually enhanced 81%, 75% and 58%, respectively, of five morphologically different abnormal masses. The circular template was extended to use data in the annulus to characterize the local background within the mammogram and to measure the local intensity within the inner disk. An adaptive, multiscale, nonparametric method for detecting abnormal masses, based on the idea of tolerance intervals is designed and implemented in the second method. The method characterizes the statistical structure of local background data by estimating the tail of the unknown local distribution. Since size, shape and intensity of abnormal masses are unknown, mammograms are processed at several different scales, pooled and combined by a rank order multiscale detector. Potential regions are clustered to form the final detected regions of interest. Empirical Free-response Receiver Operating Characteristic curves shows a Sensitivity of 1 for malignant masses in the T1b, T1c and T2 tumor categories with False Positive Rates of 4.7, 5.3 and 4.3 per image and areas of 0.90, 0.93 and 0.92 respectively. ^
Engineering, Biomedical|Health Sciences, Radiology
"Detection of masses in X-ray mammograms"
Dissertations and Master's Theses (Campus Access).