Adaptive microcalcification detection in computer aided diagnosis

Abstract

Microcalcifications are one of the earliest signs of breast cancer. In this thesis, we propose an adaptive microcalcification detection method in mammography, which gives a robust detection in CAD (computer aided diagnosis). In the proposed method, an input mammogram is pre-processed by using nonlinear image enhancement method with homomorphic filtering in wavelet domain. Adaptive denoising using wavelet shrinkage is also performed in this part. The previous contrast enhancement and denoising methods do not consider characteristics of each mammogram. This reduces the efficiency of system because each mammogram has its own properties such as noise, gray level and contrast. The proposed adaptive denoising algorithm takes the noise characteristics of each mammogram into the process. For a mammogram, its noise characteristics are obtained by analyzing the background region. The adaptive enhancement and denoising method shows better visibility of mammogram than the previous methods. Noise effect is reduced significantly while microcalcifications are more clearly seen. The enhanced mammogram is then processed in the next part of CAD system to detect microcalcification. The detection system has two stages where the first stage finds potential microcalcification pixels (ROI) and the second one detects the microcalcification within the ROIs. Both stages use artificial neural network for the purpose of detection. In the first stage, two pixel-based features, median-to-contrast and contrast-to-noise ratio, are used. The detected pixels from stage 1 are clustered to regions of interest (ROI). Four features of ROI are used for finding microcalcification from ROI. Those features are edge histogram features, high-pass masking filter features, and pixel density feature. The detected microcalcifications are grouped to form clusters of microcalcifications. Experiment is performed with three kinds of image enhancement methods: the proposed adaptive enhancement method, previ...