Research On Multiscale Denoising And Mass Segmentation Methods For Digital Mammograms

Early detection of breast cancer is a challenging task in digital mammography, since mammograms are noisy and of low contrast images. Therefore, mammograms require efficient noise suppression techniques to get reliable results. Moreover, lesion segmentation is a difficult task since lesions are embedded in and hidden by varying densities of parenchyma structures of the female breast. With the aim of providing obvious clues to radiologists for accurate diagnosis of mammograms; multiscale denoising for contrast enhancement, and robust mass segmentation approaches are studied in detail in this dissertation.First, in order to diminish noise, enhance contrast and edges of mammogram images; multiscale denoising methods are proposed based on wavelet denoising of mathematical morphology and difference of Gaussian, and then fusion at various scales. Moreover, fusion of different enhancement methods can explore distinct and sometimes complementary characteristics (such as enhancing edges and removing noise) of the given image. Furthermore, a quantitative measure, contrast improvement index (CII) is used for validation of the three proposed denoising schemes on a dataset consists of 100 mammograms. Experimental results show that the second proposed method gives average CII value 1.68 which is higher than the other denoising methods; the first method has 1.62, the third method has 1.61, VisuShrink has 0.90, BayesShrink has 0.99, PenalizedShrink has 0.92, and finally Non-Local Means (NLM) denoising has 0.69. Simulation results illustrate that the proposed multiscale denoising schemes can improve contrast and accentuate mammographic features.Second, in view of the fact that mammograms are noisy and not bimodal images, traditional segmentation methods are not robust for mass segmentation. Therefore, two mass segmentation schemes are proposed based on deformable geometric active contour model. Moreover, deformable models provide a powerful approach to the problem of accurate object segmentation; especially when the objects have no clear boundaries. Area overlap ratio (AOR) is used for validation of mass segmentation using a dataset of 160 mammographic masses. At the overlap threshold of 0.4:89%,91%, and 81% of the masses are correctly segmented with the first proposed segmentation method, the second proposed method, and the Chan-Vese (C-V) algorithm, respectively.The studied multiscale denoising approaches for contrast improvement and noise shrinking have improved the quality of denoising and contrast results. As well as, the proposed segmentation approaches of mammographic masses based on C-V geometric deformable model have increased the efficiency of mass segmentation. Hence, the subjective image quality and efficient mass segmentation have improved which are very important for human interpretation and diagnosis of digital mammograms.