Development of an automated detection system for microcalcifications lesion in mammography (in Japanese)

An automated detection system for clustered microcalcifications in digital mammograms was developed for computer-aided diagnosis (CAD) systems. We developed new detection filters of microcalcifications based on the gradient-vector image analysis. The 'triple-ring filter' extracts the region in a pattern similar to that of a microcalcification shadow. Then the 'variable-ring filter' recalculates two vector-feature values (vector-direction and vector-magnitude feature values) and determines the gray-level threshold values adaptively. These filters significantly improved the diagnostic sensitivity for the detection of microcalcifications in comparison with our previously developed scheme based on image contrast analysis. It was proved by an F (free-response) ROC examination: the true positive fraction of our new system was always 10% or more higher than our previous system while the number of false positives per image remained constant. We also developed a ''contrast-correction technique'' to correct the effects of the background content (mammary gland tissue) around the microcalcification and film contrast characteristic on its image contrast. This scheme improved the diagnostic performance for detecting the microcalcifications, and its technology was applied to the image data of two facilities where the imaging characteristic in terms of contrast was different. The diagnostic sensitivity for the detection of clustered microcalcifications in our database of 165 mammograms was 94.3% with 0.63 false positives per image, which demonstrates the effectiveness of our method. In addition, we also propose to use an artificial neural network to detect the microcalcifications in the first step instead of the triple-ring filter method: this may make it possible to develop a CAD system with an even higher detection performance level