Segmentation of microcalcifications in X-ray mammograms using entropy
thresholding

Moti Melloul

Master of Science Thesis, 2001

School of Engineering and Computer Science
The Hebrew University of Jerusalem

Abstract

Mammography is currently the best technique for reliable detection of
early, non-palpable, potentially curable breast cancer. In the past decade
there has been tredemendous interest in computer aided diagnosis (CAD) in
digital mammograms. The goal is to increase diagnosis accuracy and the
reproducibility of mammographic interpretation.

Of all breast cancer abnormalities, microcalcifications are among the most
difficult type of tumor to detect. In this thesis, we propose two new
algorithms for microcalcification segmentation in mammographic X-ray
images.

The first algorithm is a model-based algorithm which enhances mammograms
using an apriori model of calcifications. The enhancement step follows a
step of a wavelet mammogram background correction. Mammogram enhancement is
based surface fitting techniques modeling calcifications by a quadric
surface. The limitations of this algorithm are size of the calcifications
we have to know and the model fitting which fails, essentially in the case
of very small calcifications.

The second algorithm is the two-steps entropy thresholding algorithm based
on a three dimensional co-occurrence matrix. The first step removes the
slow background variations using morphological filtering. Then,
calcifications appear relatively good as bright particles on a black
background. Our goal is to obtain a binary image of white
microcalcifications over a black background after an optimal thresholding
of the filtered image. The second step of the algorithm computes an optimal
thresholding using entropy. This threshold segments calcifications from the
background.

Unlike existing methods, the entropy thresholding algorithm is fully
automatic, parameter-free, and independent of local statistics.  To test
its efficacy, we applied it to images from the Mammographic Image Analysis
Society (MIAS) database and performed a quantitative analysis of the
results with the assistance of a clinician.  We obtained detection rates of
93.75\% of true postives, 6.25% of false positives, and 2% of false
negatives.