Click here to


Are you sure ?

Yes, do it No, cancel

Impact of GPU-Accelerated Sparse-View Reconstruction for Radiation Dose Reduction in Dedicated Breast CT

H TSENG*, S Vedantham , A Karellas , University of Arizona, Tucson, AZ


(Tuesday, 7/31/2018) 4:30 PM - 5:30 PM

Room: Exhibit Hall | Forum 1

Purpose: To quantify the impact on image quality of sparse-view reconstruction techniques implemented on GPU for reduced reconstruction time, and for radiation dose reduction in cone-beam dedicated breast CT.

Methods: Full-scan (360-degree) projection datasets comprising 300 projection angles (views) from women who underwent cone-beam dedicated breast CT under an institutional review board-approved research protocol were used. Equiangular sparse view projection datasets with 150, 100, 75 and 50 views were reconstructed using six iterative
methods: maximum-likelihood estimation-maximization (MLEM), total variation based projection onto convex sets (TV-POCS), adaptive steepest descent based projection onto convex sets (ASD-POCS), simultaneous algebraic reconstruction (SART), simultaneous algebraic reconstruction with total variation (SART-TV) and prior image constrained compressed sensing (PICCS). Images were reconstructed on a GPU-enabled (GeForce TITAN XP 12GB) workstation. Signal difference to noise ratios (SDNR) between adipose and fibroglandular tissues were quantified and compared with Feldkamp-Davis-Kress (FDK) full-scan (300 views/360 degree) reconstruction.

Results: Visual analysis indicated worsening artifacts with FDK as the number of projection views were decreased. For the dataset with 50 views, none of the iterative reconstructions yielded good quality images. For the dataset with 75 views, MLEM provided reasonable quality images but with a slight degradation in resolution. Depending on the reconstruction method, projection datasets with 150 and 100 views provided good quality images. The SDNR with 100 views improved over full-scan FDK for 4 of the 6 reconstruction methods. Across all reconstruction methods, GPU accelerated reconstruction times for 100 views were 94+/-44 minutes (range: 48-168).

Conclusion: Radiation dose reduction by 50-66%, resulting in mean glandular dose of 4-5.5 mGy, is potentially attainable in dedicated breast CT using sparse-view techniques. Since screening exams are batch-read at most institutions, the reconstruction times are suitable. Optimization of each reconstruction method and reader studies are planned for future.

Funding Support, Disclosures, and Conflict of Interest: This work is supported by NIH/NCI grants R01 CA195512 and R01 CA199044. The contents are solely the responsibility of the authors and do not reflect the official views of the NIH/NCI.


Breast, Cone-beam CT, Reconstruction


IM- Breast x-ray Imaging: Breast CT

Contact Email