Improvement of clustered microcalcifications detection by using combination of reconstructed volume and projection views in digital breast tomosynthesis

Abstract

Digital Breast Tomosynthesis (DBT) is a new three-dimensional (3D) limited-angle tomography breast imaging technique that has the potential to significantly reduce camouflage effect of overlapping fibroglandular breast tissue that is a major inherent limitation for better lesion detection and identification in mammography. Since the radiologists need to interpret the large volume of DBT data, there may be chance of missing malignant lesion in breast cancer diagnosis. Thus, it is desirable to design computer-aided detection (CAD) system that aims at automatically detecting malignant lesions in DBT. Typically, for detecting clustered microcalcifications (MCs), there are two approaches. One approach uses reconstructed volume and the other one uses projection views. Each image has factors that make it difficult to detect the clustered MCs. In the reconstructed volume, MCs are distributed several slices and appear blurred. Thus, there is limitation to extract accurate features. In case of the projection views, signal to noise ratio (SNR) of MCs are low hence the sensitivity is generally lower than that of the reconstructed volume. In order to overcome aforementioned limitations, novel methods are proposed in this paper. Firstly, in order to resolve structural limitation and blur problem, maximum ray-tracing based feature extraction method is proposed that features are extracted from images that describe structural characteristics of the clustered MCs. Secondly, a novel preprocessing method in projection views is proposed. In the proposed method, MCs which are shown repeatedly in the projection views are emphasized, while random noise is suppressed. Lastly, a combined approach that fuse detection results from the reconstructed volume and projection views is devised. The combined approach is based on the fact that different approaches detect clustered MCs, while are likely to detect different false positives. Therefore, the proposed combined approach makes stronger a...