Computer aided diagnosis in radiology
Item Usage Stats
Breast cancer is one of the most deadly diseases for middle-aged women. In this thesis, computer-aided diagnosis tools are developed for the detection of breast cancer on mammograms. These tools include a detection scheme for microcalcification clusters which are an early sign of breast cancer, and a method to detect the boundaries of mass lesions. In the first microcalcification detection method we propose, a subband decomposition structure is employed. Contrary to the previous work, the detection is carried out in the subband domain. The mammogram image is first processed by a subband decomposition filter bank. The resulting subimage is analyzed to detect microcalcification clusters. In regions corresponding to the healthy breast tissue the distribution is almost Gaussian. Since microcalcifications are small, isolated bright spots, they produce outliers in the subimages and the distribution of pixels deviates from Gaussian. The subimages are divided into overlapping square regions. In each square region, skewness and kurtosis values are estimated. As third and fourth order correlation parameters, skewness and kurtosis, are measures of the asymmetry and impulsiveness of the distribution, they can be used to find the locations of microcalcification clusters. If the values of these parameters are higher than experimentally determined thresholds then the region is marked as a potential cancer area. Experimental studies indicate that this method successfully detects regions containing microcalcifications. We also propose another microcalcification detection method which uses two- dimensional (2-D) adaptive filtering and a higher order statistics based Gaussianity test. In this method, statistics of the prediction errors are computed to determine whether the samples are from a Gaussian distribution. The prediction error sequence deviates from Gaussianity around microcalcification locations because prediction of microcalcification pixels is more difficult than prediction of the pixels corresponding to healthy breast tissue. Then, we develop a new Gaussianity test which has higher sensitivity to outliers. The scheme which uses this test gives better detection performance compared to the previously proposed methods. Within the detected regions it is possible to segment individual microcalcifications. An outlier labeling and nonlinear subband decomposition based microcalcification segmentation method is also investigated. Two types of lesions, namely mass and stellate lesions, might be indicators of breast cancer. Finally, we propose a snake algorithm based scheme to detect the boundaries of mass lesions on mammograms. This scheme is compared with a recently developed region growing based boundary detection method. It is observed that the snake algorithm results in a more smooth boundary which is consistent with the morphological structure of mass lesions.