Feature extraction from mammograms by using a'priory information

Abstract

The aim of X-ray imaging is to visualize the structure of body tissue with high spatial resolution and high contrast between different kinds of tissues. Early diagnosis of breast cancer requires almost a perfect imaging system or applying sophisticated post processing algorithms to 'clean' the image and extract needed information. Diagnostic images are often assessed in clinical practice by using subjective methods which are limited by the skill of the radiologist. A numerous computer-aided diagnosis (CAD) algorithms, that assist reviewers in their decisions and reduce the numbers of false positives, have been developed. However, these systems have not been well utilized yet and there are not the consensus among scientists and practitioners which are the most effective. Straightforward deconvolution of the image, using the known imperfectness of the imaging system, is not applicable due to non-avoidable noise. The noise makes these deconvolution procedures ill-posed from a mathematical point of view i.e., the solution is not stable and must be regularized. The level of regularization depends on an a' priory information about the image formation. A regularized deconvolution procedure, based on the knowledge of the noise estimation, is proposed for de-blurring mammograms and thus enhanced the features relevant for clinical outcomes. Testing its limitations is done as well. Promising preliminary results are obtained. Further investigations are needed to apply this procedure to a real mammographic unit.