Academic Commons

Articles

Segmentation of Masses Using Continuous Scale Representations

Laine, Andrew F.; Huda, Walter; Harris, John; Chen, Dongwei

Many image processing and computer vision algorithms are surprisingly fragile when it comes to choosing various parameters and thresholds. There have been no good solutions to setting the proper value of the scale parameter in regularization networks. Local adaptation is needed in these networks to make them less brittle. Using an assumption of zero-mean, additive Gaussian noise, it can be shown that the scale parameter for edge detection problems should be proportional to the variance of any additive noise. However, these assumptions are too restrictive for most purposes, including digital mammography. More promising methods from Bayesian analysis are under investigation for setting the scale and other parameters by standard regularization-based methods. The Laplacian filter is a popular edge detector in image processing. However, knowledge of the feature size is necessary when choosing the proper scale for detection. It is common to make some assumptions about feature characteristics. For example, we can assume that microcalcifications in mammograms occur within finer scales of analysis since these features are characterized by small, high-contrast spots. Unfortunately, it is more difficult to make such general assumptions for masses in mammograms since they may assume distinct shapes and sizes, e.g., spiculated, round, or irregular. Thus, it is necessary to to identify an "optimal" scale for representation of masses. Once detected, their regions can be labeled to provide an area of local support for multiscale feature enhancement, including reinforcement of the "halo" effect, to assist in the visibility of masses. In this paper, we describe a method to segment a mass from its background using a continuous multiscale analysis. First, a suspicious region is identified, which may or may not contain a lesion. Next, the region is used as a matched filter to select a wavelet basis. A soft threshold is then applied. If the selected region contains a feature, the mammographic feature is segmented from its background for further processing.

Files

More About This Work

Academic Units
Biomedical Engineering
Published Here
August 17, 2010

Notes

Digital mammography '96 : proceedings of the 3rd International Workshop on Digital Mammography, Chicago, 9-12 June 1996 (Amsterdam ; New York : Elsevier, 1996), pp. 447-450.

Academic Commons provides global access to research and scholarship produced at Columbia University, Barnard College, Teachers College, Union Theological Seminary and Jewish Theological Seminary. Academic Commons is managed by the Columbia University Libraries.