Statistical Segmentation of Mammograms

We proposed a new algorithm for extracting abnormalities in mammograms using statistical methods. Since lesions in mammograms are disruptions of the normal patterns, it is desirable to partition the image into texture regions. Our algorithm assigns each pixel in the mammogram membership to one of a finite number of classes depending on statistical properties of a pixel and its neighbors. Both the original mammogram and its class labels are modeled as discrete parameter random fields. We estimate the pixel classes by minimizing the expected value of the number of misclassified pixels. This is known as the "maximizer of the posterior marginals" (MPM) estimate. Unlike the other MPM algorithms, our algorithm does not require that the values of all parameters of the marginal conditional probability mass functions of the pixel classes be known a priori . It combines the expectation-maximization (EM) algorithm for parameter estimation with the MPM algorithm for segmentation. In addition to segmenting abnormalities, our algorithm can also indicate the reliability of each classified pixel.

The following is a poster the authors presented at the 3rd International Workshop on Digital Mammography, 1996. To have a close look at the poster, please click on the area that interests you.

More detailed information may be found in:

Notice: the copyright to the following paper is held by the publishers. The attached PostScript file is a preprint. Please treat this material in a way consistent with the "fair use" provisions of appropriate copyright law.

M. L. Comer, S. Liu, E. J. Delp, "Statistical Segmentation of Mammograms," Proceedings of the 3nd International Workshop on Digital Mammography, June 9-12,1996, Chicago, pp. 475-478. The readme file , compressed postscript file, PDF file, and the ftp site.