(Thursday, 7/18/2019) 7:30 AM - 9:30 AM
Room: 221AB
Purpose: The development of low dose CT (LdCT) was motivated to reduce the associated risks of radiation dose. Due to heavily under-sampled data acquisition and/or low detector photon counting, the reconstruction problem is ill-posed, resulting in severe image artifacts and overwhelming noise using conventional analytical Filtered Back Projection (FBP) algorithm. Conventional sparse regularized iterative reconstruction models mitigate image artifacts and noise, but often compromise the image resolution and low-contrast object visibility. In this study, we proposed the Plug-and-Play (PnP) alternating direction method of multiplier (ADMM) framework that combines iterative reconstruction with state-of-the-art edge-preserving denoisers.
Methods: The proposed framework utilizes ADMM to solve the iterative reconstruction, formulated as a least square data fidelity term for data consistency and a non-local regularization for image smoothness. An off-the-shelf image denoiser, the Block-Matching 3D-transform shrinkage (BM3D) technique, was used to substitute for one of the ADMM modules (PnP-ADMM BM3D method). Low dose scans of the CT ACR 464 phantom and a clinical low dose lung cancer screening dataset were tested for the feasibility of the proposed PnP-ADMM BM3D framework, which was compared with FBP, Total Variation (TV), BM3D post processing on CT image, and the BM3D regularization.
Results: Compared with FBP reconstruction of higher dose scans, the PnP-ADMM BM3D significantly reduced image noise and resolved an equivalent or increased level of image detail. With a comparable level of image noise, the proposed PnP-ADMM BM3D distinguished the 9 lp/cm line pairs of the ACR phantom and lobar fissures in the lung screening image data, while these fine details were not clearly visible in other reconstruction methods, including conventional iterative methods and post processing (denoising) methods.
Conclusion: The flexible PnP ADMM framework combines CT iterative reconstruction with state-of-the-art edge-preserving denoisers, achieving reduced noise, improved contrast, and maintains high resolution performance for low-dose CT.
Funding Support, Disclosures, and Conflict of Interest: NIH R01CA230278 NIH R44CA183390 NIH R01CA188300
Keywords
CT, Reconstruction, Optimization
Taxonomy
Contact Email
