Room: Exhibit Hall | Forum 1
Purpose: In digital breast tomosynthesis (DBT) reconstruction, back-projection (BP) method canâ€™t produce images with high contrast features and sharp edges and costing a long processing time is a common disadvantage of traditional iterative methods. In this work, we developed a fast alternating direction method of multipliers (ADMM) and multiple graphics processing units (GPUs)-based algorithm to reconstruct high quality DBT image efficiently
Methods: The DBT image is reconstructed by minimizing an energy functional consisting of a data fidelity term and a total variation (TV) regularization term via ADMM and based on GPUs. The algorithm mainly includes five steps: 1) Distribute the subsets of projections to each GPU; 2) Perform CGLS to obtain preliminary images at all GPUs parallelly; 3) Conduct reduction among GPUs and combine preliminary images; 4) Perform shrinkage operator to obtain final image; 5) Broadcast reconstruction image to all GPUs; 6) Repeat step 2)~5) until our criteria is met to terminate the iteration
Results: We test our algorithm on a simulated digital phantom and a physical phantom, CIRS-011A. The the mean square error (MSE) and peak signal noise ratio (PSNR) were employed as our evaluation metrics of the reconstructed images. Phantom studies indicate that our algorithm resulted in lower MSE and higher PSNR values than the BP and MLEM methods. In the images reconstructed by our method, the features are much more conspicuous with sharper edges and higher contrast. In terms of reconstruction efficiency, the computation time of our method with two GPU cards is half shorter than that of MLEM method.
Conclusion: The developed reconstruction algorithm based on ADMM and GPUs could effectively reconstruct the high quality DBT image. The high computational efficiency makes this iterative DBT reconstruction approach feasible in clinical environments.
Funding Support, Disclosures, and Conflict of Interest: National Natural Science Foundation of China (81301940 and 81428019), National Key Research and Development Program (2016YFA0202003), Guangdong Natural Science Foundation of China (2016A030310388 and 2017A030313692), and Southern Medical University Startup fund (LX2016N003).