Pectoral muscle segmentation in mammograms based on cartoon-texture decomposition

Pectoral muscle segmentation on medio-lateral oblique views of mammograms represents an important preprocessing step in many mammographic image analysis tasks. Although its location can be perceptually obvious for a human observer, the variability in shape, size, and intensities of the pectoral muscle boundary turns its automatic segmentation into a challenging problem. In this work we propose to decompose the input mammogram into its textural and structural components at different scales prior to dynamically thresholding it into several levels. The resulting segmentations are refined with an active contour model and merged together by means of a simple voting scheme to remove possible outliers. Our method performs well compared to several other state-ofthe- art techniques. An average DICE similarity coefficient of 0.91 and mean Hausdorff distance of 3.66 ± 3.23 mm. validate our approach.