Automatic Detection of Clustered Microcalcifications Using Morphological Reconstruction

Abstract

This paper presents a method for automatic detection of clusters of microcalcifications (MCCs) in mammograms. As MCCs are small regions made up by groups of pixels brighter than their neighbours, in order to detect them the operators have to be based on the size and contrast. In mammography linear operators have been used like Laplacian-of-a-Gaussian (LoG), among others, approximated for the discrete integer approximation with the Difference-of-Gaussian (DoG) filter, as proposed in [1]. In this paper we explore operators based on non-linear theory, specifically based on mathematical morphology. This discipline has already been used in previous work [2] by means of the watershed, but in that paper the process needs user-introduced markers to reduce the oversegmentation produced by the watershed, so it is not a fully automatic detection method. In a previous work [3] we proposed to use the residue of the supremum of a number of openings each of them with a linear structuring element jointly with the Markov random field model (MRF) proposed in [4], which can eliminate the major part of the connective tissue reducing the false positive rate. In some cases this morphological operator is not enough. In this paper we studied the contribution of morphological reconstruction to improve the elimination of the connective tissue and so to reduce the false positive rate.