Total Variation Regularization in Digital Breast Tomosynthesis: Regularization Parameter Determination based on Small Structures Segmentation Rates

Regularization approaches for the limited-angle reconstruction problem in digital breast tomosynthesis are widelyused. Though, their benefits depend largely upon a suitable regularization parameter estimation. We aim to evaluate the reconstruction quality of precise small contrast features objectively with the help of an automated process. These features were represented by so-called Landolt ring (LR) structures of descending sizes contained in an especially designed mammography test object (Quart Mam/Digi Phantom). A GPU-based iterative Barzilai-Borwein (BB) algorithm is applied to solve the inverse reconstruction problem using total variation (TV) regularization. Exemplarily, we analyzed limitedangle breast projection images from a commercially available digital breast tomosynthesis (DBT) system (Siemens Mammomat Inspiration). We show that the TV regularization parameter and number of iterations can be chosen in such a way that the detection rate for the LR features is considerably higher than that obtained from the manufacturer’s reconstruction (modified filtered backprojection).