A novel approach for the mammary gland region segmentation of Breast Ultrasound exams is proposed. This method is important because the mammary gland is the Region of Interest for pathological diagnosis. Five different pre-processing methods that enhance the transition areas or remove the speckle of the ultrasound images were selected: Non-linear diffusion, Speckle Reducing Anisotropic Diffusion, Entropy filter, Laplacian filter and Homomorphic filter. The results of these processing methods define the features that are used as descriptors for a K-Means and SVM classifier or as weak classifiers by an Adaboost classifier. The pixel classification results in a rough tissue segmentation. A new method is proposed to interpolate the classification results into an accurate tissue separation line, using graph theory. This step overcomes the problem of the discontinuities between the different classified areas. The developed segmentation method was applied to a database with 61 images, 34 without masses and 27 with masses collected using digital support, and segmented by an experienced medical oncologist in Centro Hospitalar da Cova da Beira in Portugal. The presented results were obtained using cross-validation.

Date Published: 13 March 2013
PDF: 9 pages
Proc. SPIE 8669, Medical Imaging 2013: Image Processing, 866948 (13 March 2013); doi: 10.1117/12.2006967

Published in SPIE Proceedings Vol. 8669:
Medical Imaging 2013: Image Processing
Sebastien Ourselin; David R. Haynor, Editor(s)