Room: Room 202
Purpose: In digital breast tomosynthesis (DBT), the conspicuity of fiber-like signals exhibits in-plane orientation dependence. The purpose of this study is to investigate the design of reconstruction algorithms to mitigate this effect.
Methods: Four image reconstruction algorithms are investigated: filtered backprojection (FBP) with Hanning window apodization, a modified implementation of FBP employing a boost to the low-frequency response of the ramp filter (mFBP), Tikhonov penalized least squares (T-PLSQ), and roughness penalized least squares (R-PLSQ). Using a Hologic Selenia Dimensions scanner, data is acquired from an ACR digital mammography accreditation phantom. Reconstructions of the phantom are used to compare fiber-like signal conspicuity between algorithms. The width of the Hanning apodizing window for FBP reconstruction is chosen based on subjective visualization, and the regularization strengths employed in the other three algorithms are chosen based on matching this reference reconstruction's depth resolution. A simulation study is performed to quantify the depth resolution for each algorithm using the artifact spread function of a point-like signal.
Results: Orientation dependent conspicuity is observed in each of the four algorithms at low regularization strengths, with fibers oriented parallel to the detector and the plane containing the source-trajectory being less conspicuous than their orthogonal counterparts. With increasing regularization strength, this orientation dependence is reduced. The regularization parameter is thereby seen to control an orientation-dependence/depth-resolution tradeoff. This tradeoff is found to be more favorable for mFBP, T-PLSQ, and R-PLSQ than for FBP with Hanning window apodization.
Conclusion: The regularization strength employed in DBT image reconstruction not only controls noise level and depth resolution, but also the orientation dependence of fiber-like signal conspicuity. Three alternative algorithm designs are seen to exhibit a better tradeoff between orientation dependence and depth resolution than FBP with Hanning window apodization.