Room: Exhibit Hall | Forum 1
Purpose: To enable low dose CT using filtered backprojection (FBP) reconstruction, three major classes of denoising schemes have been proposed: (i) denoising in the raw counts domain before the log-transformation step, (ii) denoising line integral data in the log-transform domain, and (iii) image domain denoising. Various authors have found success and argued for operating within each domain, yet a clear consensus has not been reached. The purpose of this study was to investigate the domain that provides the best overall imaging performance.
Methods: To assess the imaging performance of denoising in each domain an anisotropic scenario that includes high overall noise magnitude and noise streaks was considered. The shoulder region of an anthropomorphic phantom was scanned at 1.9 and 0.5 mGy using a benchtop CBCT imaging system. Fifty repeated scans were performed at each dose level to enable accurate local frequency-dependent measurements of noise and spatial resolution (2D NPS and MTF). Noise variance and MTFâ‚?â‚€ were used as 0D surrogates of noise and resolution. Additionally, a scalar â€œisotropyâ€? metric of NPS and MTF was defined. This metric ranges from 0-1: e.g. for the NPS, 0 means that noise is highly directional (streaky), whereas 1 means it is perfectly isotropic. Preliminary results were obtained with a shift-invariant boxcar filter (box size 1-7).
Results: Quantitative assessment demonstrated three major
results: (1) noise streaks are more effectively reduced when denoising in the raw counts domain; (2) noise variance and NPS isotropy scores show that raw counts domain denoising provides lower overall noise magnitude and higher isotropy scores (higher noise streaks reduction) than the other two methods; (3) the spatial resolution results showed no significant difference across the methods included in this study.
Conclusion: Performing denoising in the raw counts domain provides better overall imaging performance compared to log-transform domain and image domain denoising.
Not Applicable / None Entered.