Integration of Low-Level Processing to Facilitate Microcalcification Detection

Abstract

The performance of microcalcification detection algorithms is currently not good enough for them to be used in a clinical setting. Most attempts to improve their performance consist of devising increasingly smarter high-level detection schemes. In contrast, we believe that application of low-level model-based image processing can reduce the number of false positives generated by existing detection algorithms. In this paper, we show how to identify those pixels which tend to be systematically labeled falsely as microcalcifications because of their similarity in radiological appearance to microcalcifications, namely screen-film ‘shot’ noise. In one of the most successful algorithms, Karssemeijer [4] treats such noise at the segmentation step, not in his preprocessing step which aims at making noise spectrally flat by rescaling pixel values, by defining an extra label class in his Markov random field (MRF) model.