Hier finden Sie wissenschaftliche Publikationen aus den Fraunhofer-Instituten.

Computer-aided detection of lesions in digital breast tomosynthesis images

: Prinzen, Martin; Wagner, Florian; Nowack, Sebastian; Schulz-Wendtland, Rüdiger; Paulus, Dietrich; Wittenberg, Thomas


Deserno, T.M. ; Deutsches Krebsforschungszentrum, Abteilung für Medizinische und Biologische Informatik; Berufsverband Medizinischer Informatiker -BVMI-; Deutsche Gesellschaft für Computer- und Roboterassistierte Chirurgie -CURAC-:
Bildverarbeitung für die Medizin 2014 : Algorithmen - Systeme - Anwendungen; Proceedings des Workshops vom 16. bis 18. März 2014 in Aachen
Berlin: Springer, 2014 (Informatik aktuell)
ISBN: 3-642-54110-0
ISBN: 978-3-642-54110-0 (Print)
ISBN: 978-3-642-54111-7 (Online)
Workshop Bildverarbeitung für die Medizin (BVM) <2014, Aachen>
Fraunhofer IIS ()
Texturmerkmal; Mammographie; Klassifikation; Kantendetektion; Defekterkennung; 3D Bildverarbeitung

The most common cancer among women in the western world is breast cancer. Early detection of lesions greatly influences the progress and success of its treatment. Digital breast tomosynthesis (DBT) is a new imaging technique that facilitates a three-dimensional reconstruction of the breast. DBT reduces superimposition of breast tissues and provides better insight into the breast compared to the common digital mammography. In order to assist radiologists with the examination and assessment of the large amount of DBT data, a computer aided detection (CADe) of focal lesions can be an essential tool, leading to increased sensitivity and specificity. We present and compare two different approaches for a fully automated detection of lesions in DBT data using voxel-wise classification, one being the state of the art and the other one an enhancement. Multiple difference of Gaussians detect lesions based on their common higher intensity and contrast in relation to surrounding tissue. A gradient orientation analysis detects round and spiculated lesions, even when they are weak in contrast and intensity. By combining these features and using a support vector machine, a classification performance of 88% can be achieved.