Heang-Ping Chan PhD, Presenter: Nothing to Disclose

Jun Wei PhD, Abstract Co-Author: Nothing to Disclose

Berkman Sahiner PhD, Abstract Co-Author: Nothing to Disclose

Elizabeth Ann Rafferty MD, Abstract Co-Author: Nothing to Disclose

Tao Wu PhD, Abstract Co-Author: Nothing to Disclose

Jun Ge PhD, Abstract Co-Author: Nothing to Disclose

Marilyn A. Roubidoux MD, Abstract Co-Author: Nothing to Disclose

Richard H. Moore MD, PhD, Abstract Co-Author: Nothing to Disclose

Daniel Benjamin Kopans MD, Abstract Co-Author: Nothing to Disclose

Lubomir M. Hadjiiski PhD, Abstract Co-Author: Nothing to Disclose

Mark Alan Helvie MD, Abstract Co-Author: Nothing to Disclose

et al, Abstract Co-Author: Nothing to Disclose

To compare computerized mass detection on the reconstructed digital breast tomosynthesis (DBT) mammograms and on the unreconstructed projection views (PVs).

DBT is a new modality that holds the promise of improving breast cancer detection. We are developing computer-aided diagnosis methods for detection of breast lesions on DBT mammograms. In this preliminary study, we used a data set of 26 DBT cases acquired by a GE DBT prototype system at the Breast Imaging Research Lab of Massachusetts General Hospital. The DBT system acquired 11 PVs of the compressed breast over a 50-degree arc in the MLO view. The total dose for the 11 PVs was less than 1.5 times that of a single standard film. DBT slices were reconstructed at 1-mm slice spacing using an iterative maximum-likelihood algorithm. The cases included 23 masses (13 malignant) and 3 areas of architectural distortion (2 malignant). The DBT slices per case ranged from 37 to 89 (mean=60.1). We evaluated two approaches: the first using DBT slices as input and the second directly processing the 11 PVs. For DBT slices, gradient field analysis was applied to both the 2D slices and the 3D volume for prescreening of lesion candidates. For PVs, the raw images were first preprocessed with a Laplacian pyramid multi-scale enhancement scheme and prescreening was applied to each enhanced PV. In each approach, morphologic and texture features were extracted from the lesion candidates and classifiers were developed to reduce false positives (FPs). In the final step, a voting scheme was designed to estimate the likelihood that a detected object was a true lesion based on the multiple-image information. Two-fold cross-validation was used for training and testing the algorithms and the two approaches were compared by using FROC analysis.

For testing, at 92% sensitivity, the FP rate was 2.8/case for DBT slices and 3.7/case for PVs. The FP rate decreased to 2.5 for DBT slices and 2.8 for PVs at 80% sensitivity.

The detection accuracy on DBT slices is higher than that on PVs, probably because of the reduced camouflaging effects of overlapping tissue. Further work is underway to improve the detection algorithms and to enlarge the data set.

Chan, H, Wei, J, Sahiner, B, Rafferty, E, Wu, T, Ge, J, Roubidoux, M, Moore, R, Kopans, D, Hadjiiski, L, Helvie, M, et al, , Computer-aided Detection on Digital Breast Tomosynthesis (DBT) Mammograms: Comparison of Two Approaches.  Radiological Society of North America 2004 Scientific Assembly and Annual Meeting, November 28 - December 3, 2004 ,Chicago IL. http://archive.rsna.org/2004/4411939.html