A multi-scale 3D radial gradient filter for computerized mass detection in digital tomosynthesis breast images

• Published in: International Congress series Vol. 1281; pp. 1058 - 1062
• Main Authors: Reiser, I., Nishikawa, R.M., Giger, M.L., Kopans, D.B., Rafferty, E.A., Wu, T., Moore, R.
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01-05-2005
• Subjects: Computer-aided detection; Digital mammography; Tomosynthesis; Digital mammography; Computer-aided detection; Tomosynthesis
• ISSN: 0531-5131; 1873-6157
• DOI: 10.1016/j.ics.2005.03.171

We have developed a pre-selection algorithm to identify mass lesion candidates in digital breast tomosynthesis (DBT) images. This algorithm is designed to achieve computationally efficient pre-selection of lesion candidate locations in the reconstructed breast volume. First, the breast volume is filtered with a radial gradient filter to enhance spherical structures. The filter is designed such that the convolution can be carried out in the spatial frequency domain. Variation of mass lesion size is accounted for by varying the scale of the radial gradient filter. For the subsequent analysis, the dimensionality of the feature volume is reduced through maximum intensity projection, without loss of information for the present database. The database consisted of 36 breast volumes, 21 of which had mass lesions. DBT volumes were reconstructed from 11 projections using a ML-EM algorithm. For this database, we found a sensitivity of 100% at 23 false positives and 90% sensitivity at 13 false positives. False positives can be reduced through subsequent feature analysis.