Segmentation of mammographic images for computer aided diagnosis

Abstract

Computer-aided diagnosis (CAD) systems are currently at the heart of many clinical protocols since they signi cantly improve diagnosis and therefore medical care. However, designing a CAD tool for early breast cancer detection remains a di cult and challenging task. In fact, it is hard to conceptualize an expert radiologist's judgment. This research work therefore puts forward a hierarchical architecture for the design of a robust and e cient CAD tool for breast cancer detection. More precisely, it focuses on the reduction of false alarms rate through the identi cation of image regions of foremost interest (dense breast tissues). Adapted strategies for breast cancer pattern identification can then be applied in priority. The approach hereby introduced relies on two macro-steps. Firstly, raw mammographic images are gotten rid of poorly informative image regions (background and muscle tissues) impairing automatic breast tissue analysis and cancer signs identi cation. Then, a more advanced analysis is performed on the remaining image to characterize dense breast tissues with respect to their density in order to identify potential cancerous areas. This PhD manuscript starts with useful insights into mammograms followed by a number of image processing developments to carry out the two macro-steps mentioned above. In the first macro-step, the dynamic range of gray level intensities in dark regions is stretched to enhance the contrast between tissues and background. This enhancement process favors accurate breast region extraction and suppression of all unwanted patterns in the background image region. A second segmentation follows background suppression. Indeed, some muscle tissues regularly tampering breast tissue analysis remains inlaid in the foreground region i.e pectoral muscle tissues. Extracting pectoral muscle tissues is both hard and challenging due to mammogram peculiarities such as the overlap between dense breast and pectoral muscle tissues. In such conditions, even exploiting spatial information during the clustering process of the fuzzy C-means algorithm does not always produce a relevant segmentation of this image region. To overcome this difficulty, a new validation process followed by a refinement strategy are proposed to detect and correct the segmentation imperfections and thus enabling accurate pectoral muscle region extraction. The second macro-step is devoted to breast tissue density analysis. To address the variability issues observed in gray levels distributions with respect to mammographic density classes, we introduce an optimized gray level transport map for mammographic image contrast standardization. Despite the lack of a target histogram distribution, useful parameters can be derived allowing an easier discrimination of mammogram density classes. Thanks to this technique, dense tissues regions are segmented using simple thresholding. We prove that the dense region areas computed from segmented images are highly correlated to density classes from an annotated dataset.