Analysis of mammographic microcalcifications using gray-level image structure features
IEEE Transactions on Medical Imaging
Recent studies indicate that 1 in 14 Australian women will develop breast cancer with an annual mortality rate approaching 3000. While there appears to be little prospect for preventing breast cancer, the introduction of routine mammographic screening programmes has been successful in reducing mortality. However, these screening programmes present the examining radiologist with an increased case load, increasing the chance of improper diagnosis. Computer-Aided Diagnosis (CAD) schemes using
... schemes using digital image processing techniques attempt to improve the detection performance and efficiency of mammography screening. These computerised systems relieve the radiologist from inspecting large numbers of mammograms, and allow the detailed consideration of the more suspicious cases. Microcalcification clusters have been specifically targeted as a reliable early indicator of breast cancer. Microcalcifications are tiny granule-like deposits of calcium. Their presence is associated with a higher probability of cancerous regions. The identification of such regions is a difficult process because of the small size, the weak definition, and the low contrast. Moreover, in addition to these microcalcifications, a mammogram preserves considerable information about the background tissue, including anatomical structures such as ducts and glands. Such information, however, is not useful to the radiologist. Recent advances in digital image acquisition and the decreasing cost-performance ratio of computers have made semiautomated mammogram analysis techniques more viable. Research in mammogram imaging has concentrated on locating suspicious regions such as microcalcification clusters. Different features extracted from mammographic images have been used to locate these clusters. Davies and Dance used shape features such as area, shape parameters, edge strength and clustering as features. Dhawan et al. recommend greylevel features such as histogram skewness, eigenmass, energy, entropy, and moments being better for low-contrast images . Classification of these features is a critical step in diagnosis. Several classification techniques have been developed including clustering methods, pattern recognition, expert systems and neural networks. Each has its own merits and suitability for discriminating certain features, however their efficacy is completely dependent of the quality of the extracted feature set. Our work expands upon the results of Dhawan et al.  and investigates the performance of several features extracted from a Multiscale Analysis of mammographic images using the Wavelet transform. The Wavelet transform decomposes a given image at several resolution levels. Some of the features mentioned above may or may not appear at all resolution levels of a mammographic image. The features extracted from the different levels of these images are used in a statistical classifier for discrimination of benign and malignant microcalcification clusters. Initial results using Bayesian classifiers indicate that a Multiscale Analysis preceeding classification can better represent the local variations important for the detection of microcalcifications compared to statistical and shape features extracted only from the original image (scale 0).