Improving Image Reconstruction for Digital Breast Tomosynthesis

Abstract

Digital breast tomosynthesis (DBT) has been developed to reduce the issue of overlapping tissue in conventional 2-D mammography for breast cancer screening and diagnosis. In the DBT procedure, the patient’s breast is compressed with a paddle and a sequence of x-ray projections is taken within a small angular range. Tomographic reconstruction algorithms are then applied to these projections, generating tomosynthesized image slices of the breast, such that radiologists can read the breast slice by slice. Studies have shown that DBT can reduce both false-negative diagnoses of breast cancer and false-positive recalls compared to mammography alone.

This dissertation focuses on improving image quality for DBT reconstruction. Chapter I briefly introduces the concept of DBT and the inspiration of my study. Chapter II covers the background of my research including the concept of image reconstruction, the geometry of our experimental DBT system and figures of merit for image quality. Chapter III introduces our study of the segmented separable footprint (SG) projector. By taking into account the finite size of detector element, the SG projector improves the accuracy of forward projections in iterative image reconstruction. Due to the more efficient access to memory, the SG projector is also faster than the traditional ray-tracing (RT) projector. We applied the SG projector to regular and subpixel reconstructions and demonstrated its effectiveness. Chapter IV introduces a new DBT reconstruction method with detector blur and correlated noise modeling, called the SQS-DBCN algorithm. The SQS-DBCN algorithm is able to significantly enhance microcalcifications (MC) in DBT while preserving the appearance of the soft tissue and mass margin. Comparisons between the SQS-DBCN algorithm and several modified versions of the SQS-DBCN algorithm indicate the importance of modeling different components of the system physics at the same time.

Chapter V investigates truncated projection artifact (TPA) removal algorithms. Among the three algorithms we proposed, the pre-reconstruction-based projection view (PV) extrapolation method provides the best performance. Possible improvements of the other two TPA removal algorithms have been discussed. Chapter VI of this dissertation examines the effect of source blur on DBT reconstruction. Our analytical calculation demonstrates that the point spread function (PSF) of source blur is highly shift-variant. We used CatSim to simulate digital phantoms. Analysis on the reconstructed images demonstrates that a typical finite-sized focal spot (~ 0.3 mm) will not affect the image quality if the x-ray tube is stationary during the data acquisition. For DBT systems with continuous-motion data acquisition, the motion of the x-ray tube is the main cause of the effective source blur and will cause loss in the contrast of objects. Therefore modeling the source blur for these DBT systems could potentially improve the reconstructed image quality. The final chapter of this dissertation discusses a few future studies that are inspired by my PhD research.