Fast Low-Dose CT Image Processing Using Improved Parallelized Nonlocal Means Filtering

Abstract

International audienceAlthough effectively reducing the radiation exposure to patients, low dose CT (LDCT) images are often significantly degraded by severely increased mottled noise/artifacts, which can lead to lowered diagnostic accuracy in clinic. The nonlocal means (NLM) filtering can effectively remove mottled noise/artifacts by utilizing large-scale patch similarity information in LDCT images. But the NLM filtering application in LDCT imaging is also accompanied with high computation cost as a large searching window is often required to include much neighboring information for noise/artifact suppression. To accelerate the NLM filtering and improve its clinical feasibility, we propose in this paper an improved GPUbased parallelization approach. In addition to the straight pixel wise parallelization, the improved parallelization approach exploits the high I/O speed of GPU shared memory. Quantitative experiment demonstrates that significant acceleration is achieved with respect to the traditional pixel-wise parallelizatio

Applications of nonlocal means algorithm in low‐dose X‐ray CT image processing and reconstruction: A review

“…One time‐consuming procedure is the calculation of patch distance within a large SW. Various schemes have been proposed to speed up this process, including preselecting patches, neglecting Gaussian kernel, taking advantage of symmetry, and employing GPU for parallel processing . For NLM‐regularized image reconstruction, employing an iterative algorithm to optimize the objective function is also time‐consuming because of multiple re‐projection and back‐projection operation cycles in the projection and image domains.…”

Optimized Parallelization for Nonlocal Means Based Low Dose CT Image Processing

“…For GPU with fast single-precision floating processing, the main computation cost of ( 10 ) lies in the data accessing operation of the global memory because the time cost in shared memory accessing is trivial when compared to global memory accessing. The computational complexity of ( 11 ) can be roughly estimated to be O (2 B + 1) [ 21 ].…”