作者: Libo Zhang , Benqiang Yang , Zhikun Zhuang , Yining Hu , Yang Chen
DOI: 10.1155/2015/790313
关键词: Acceleration 、 Shared memory 、 Parallel computing 、 Artificial intelligence 、 Similarity (geometry) 、 Signal-to-noise ratio 、 Low dose ct 、 Image processing 、 Computation 、 Noise 、 Computer vision 、 Computer science
摘要: 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 effectively remove noise/artifacts utilizing large-scale patch similarity information LDCT images. But the NLM application imaging also requires high computation cost because intensive calculation within a large searching window is required be used include enough structure-similarity for noise/artifact suppression. To improve its clinical feasibility, this study we further optimize parallelization of avoiding repeated with row-wise intensity and symmetry weight calculation. shared memory fast I/O speed proposed method. Quantitative experiment demonstrates that significant acceleration achieved respect traditional straight pixel-wise parallelization.