Ingrid Reiser PhD, Presenter: Nothing to Disclose

Robert Mark Nishikawa PhD, Abstract Co-Author: Shareholder, Hologic, Inc, Bedford, MA Royalties, Hologic, Inc, Bedford, MA Research funded, Hologic, Inc, Bedford, MA Consultant, Fuji Photo Film Co, Ltd, Stamford, CT Scientific Advisory Board, Dexela Limited, United Kingdom

Helene Derand PhD, Abstract Co-Author: Employee, XCounter AB

Christer Ullberg MSc, Abstract Co-Author: Employee, XCounter AB

Tom Francke PhD, Abstract Co-Author: Employee, XCounter AB

Karin Lindman, Abstract Co-Author: Employee, XCounter AB

Daniel B. Kopans MD, Abstract Co-Author: Research support, General Electric Company

Richard H. Moore MD, PhD, Abstract Co-Author: Research support, General Electric Company

et al, Abstract Co-Author: Nothing to Disclose

et al, Abstract Co-Author: Nothing to Disclose

The purpose of this study was to investigate whether a mass detection algorithm can be applied to tomosynthesis breast images acquired with different acquisition parameters, such as scan angle and number of projection views, and dose.

Two databases of patient images were used in this study. The first set of 82 breast volumes containing mass lesions was acquired at the Massachusetts General Hospital on the first GE tomosynthesis prototype which acquired 11 projection views over a 50 degrees scan. Detector pixel size was 100 microns. The second database contained 9 mass lesions and 5 mass/AD. Those cases were acquired on the XCounter tomosynthesis system which acquires 48 projection views over a scan angle of 22 degrees. Detector pixel size was 60 microns. Dataset 1 was reconstructed using iterative ML-EM, while the second dataset was reconstructed with a proprietary algorithm (iterative). An automated mass detection algorithm has been developed previously on a subset of 21 cases from the first database. The initial detection stage of this algorithm consists of a 3D radial gradient filter, which computes the average radial gradient within a shell at multiple scales. Filtered images at multiple scales are combined through maximum rule. This lesion detection filter was applied to both databases and evaluated.

Preliminary results indicate that algorithm performance on both datasets is similar (database1: 83% sensitivity at 24 false positives per breast volume; database2: 86%sensitivity at 15 false postives per volume). In both datasets, lesions with ill-defined margins tended to be missed.

The results of this pilot study indicate that computerized detection algorithms can be robust across tomosynthesis images acquired with different acquisition geometries.

Automated mass detection for digital breast tomosynthesis may provide an aid to radiologists in detecting breast lesions. Ideally, CADe techniques should work on images collected under different conditions. Our initial results indicate that our approach may be robust.

Reiser, I, Nishikawa, R, Derand, H, Ullberg, C, Francke, T, Lindman, K, Kopans, D, Moore, R, et al, , et al, , Robustness of an Algorithm for Computerized Mass Detection in Digital Breast Tomosynthesis.  Radiological Society of North America 2007 Scientific Assembly and Annual Meeting, November 25 - November 30, 2007 ,Chicago IL. http://archive.rsna.org/2007/5016285.html