Features extraction for a precise characterization of microcalcifications in mammograms

Microcalcifications are an important sign for breast cancer diagnosis. Here the authors propose a three steps approach for microcalcifications detection and characterization. Firstly, a new and efficient non-linear filter, based on the global prior of the microcalcifications´ shape, is presented. This filter limits the false detections due to noise while providing seeds representative of suspicious regions. In a second step a precise segmentation of the suspicious regions is provided by a region growing technique initialized from the previously detected seeds. In a last step, the individual potential microcalcifications are characterized by a set of features and grouped in clusters. This step separates the false detections from the true ones, on the basis of the high level prior of the microcalcifications, and provides quantitative information useful for the physicians´ diagnosis. This global approach has been tested and evaluated on mammograms from the MIAS database, representative of different pathologies and breast tissue structures