Room: AAPM ePoster Library
Purpose: Digital breast tomosynthesis (DBT) systems have a wide variety of detector types, acquisition geometries and reconstruction techniques. Because the choice of reconstruction affects image quality, it is useful to decouple the reconstruction during optimization of system parameters. A channelized multi-projection observer is developed that does not require a reconstructor but uses the full set of projections for characterization of DBT. A partial least squares (PLS) estimation of the channels is used for detecting lesions of complex shape.
Methods: A 3D breast volume with anatomic noise background was simulated using a binarized 3D voxel distribution with power-law spectral content. Lesion inserts of either spherical or stellate lesion shapes were used. Simulated projections from DBT system were created for both signal-present and signal-absent cases for up to 3000 unique volumes suitable for alternative forced choice experiments. The PLS estimated channels were extracted from a quantum-noise free test set. Using a decoupled covariance matrix estimation approach, the detectability metric, d', is calculated for both lesion shapes across a range of doses and electronic noise. An equal dose per projection scheme was compared to one in which half the dose is delivered to the central projection.
Results: The PLS estimation improved with number of volumes used in training, although the improvement was minor for n>1000 volumes. The d' improved with increasing channel number, asymptotically beyond 10 channels. The stellate lesion shows reduced detectability at low angular ranges compared to spherical lesions. The central dose scheme showed 40% increased d' performance under low electronic noise conditions, but not at higher doses (>1 mGy).
Conclusion: A channelized multi-projection observer was developed for use in evaluating virtual models of DBT systems without the need for a reconstruction step. This can be a useful tool in optimizing and evaluation the performance of different acquisition approaches.
Funding Support, Disclosures, and Conflict of Interest: Our lab has a research collaboration with GE Healthcare on various topics in digital mammography and tomosynthesis