Program Information
Task-Based Parameter Selection for Linear Iterative Image Reconstruction in Breast Tomosynthesis
S Rose*, A Sanchez , I Reiser , E Sidky , X Pan , The University of Chicago, Chicago, IL
Presentations
WE-G-601-3 (Wednesday, August 2, 2017) 4:30 PM - 6:00 PM Room: 601
Purpose: To compare two simulation, task-based image quality metrics assessing the effect of regularization strength on microcalcification detectability in DBT reconstruction.
Methods: Two task-based image quality metrics are investigated: a region-of-interest (ROI) Hotelling observer (HO) for a signal-known-exactly/background-known-exactly detection task and an ROI non-prewhitening (NPW) observer. Simulation studies are performed in which regularization strength is varied for identity (LSQI) and derivative (LSQD) Tikhonov regularized least-squares reconstruction. The metrics are applied to the task of microcalcification detection, modeled using a 0.32mm high-contrast signal. The metrics are calculated in closed form, not requiring noise realizations, to facilitate efficient investigation of parameter spaces involved in reconstruction. Trends in the ROI-HO and ROI-NPW metrics are compared with 3D reconstructions from ACR mammography accreditation phantom data acquired with a Hologic Selenia Dimensions DBT system.
Results: The efficiencies --- squared ratios of signal-to-noise ratio (SNR) in the image to SNR of the HO applied to the data --- of the ROI-HO increase with increasing regularization strength, eventually saturating for both algorithms. For LSQI, the ROI-NPW exhibits similar behavior while for LSQD it exhibits a peak efficiency at intermediate regularization strength followed by a decrease in efficiency. In ACR reconstructions, the reconstructed specks appear more conspicuous as the noise level is reduced by increasing regularization strength.
Conclusion: The efficiency curves for the ROI-HO suggest that information relevant to task performance is better preserved with increasing regularization but also suggest a point of diminishing returns. For LSQD reconstruction, the ROI-NPW efficiency exhibits a peak while the ROI-HO does not, demonstrating that pre-whitening affects trends in performance of the task. The trend of increasing conspicuity with increasing regularization in the ACR data reconstructions appears to coincide with the ROI-HO and ROI-NPW efficiency trends for LSQI and the ROI-HO trends for LSQD reconstruction.
Contact Email: