To compare the noise-magnitude and noise-texture obtained using strong kernel across two generations of iterative reconstruction (IR) algorithms proposed by three manufacturers.

Five computed tomography (CT) systems equipped with two generations of IR algorithm (hybrid/statistical IR [H/SIR] or full/partial model-based IR [MBIR]) were compared. Acquisitions on Catphan 600 phantom were performed at 120kV and three dose levels (CTDI_vol: 3, 7 and 12mGy). Raw data were reconstructed using standard “bone” kernel for filtered back projection and one iterative level of two generations of IR algorithms. Contrast-to-noise ratio (CNR) was computed using three regions of interest placed semi-automatically: two placed in the low-density polyethylene and Teflon inserts and another placed on the solid water. Noise power spectrum (NPS) was computed to assess the NPS-peak and noise-texture.

CNR was significantly greater in MBIR compared to H/SIR algorithms for all CT systems (P<0.01). CNR were improved on average from H/SIR to MBIR of 36±14% [SD] (range: 24–57%) for GE-Healthcare, 109±19 [SD] % (range: 89–139%) for Philips Healthcare and 42±5 [SD] % (range: 36–47%) for Siemens Healthineers. The mean NPS peak decreased from H/SIR to MBIR by −41±6 [SD] % (range: −47–−35%) for GE Healthcare, −79±3 [SD] % (range: −82–−76%) for Philips Healthcare and −52±2 [SD] % (range: −54–−51%) for Siemens Healthineers systems. NPS spatial frequencies were greater with MBIR than with H/SIR for Philips Healthcare (20 ± 2 [SD] %; range: 19–22%) and for Siemens Healthineers (9±5 [SD] %; range: 4–14%) but lower for GE Healthcare (−17±3 [SD] %; range: −14–−20%).

Using bone kernel with recent MBIR algorithms reduces the noise-magnitude for all CT systems assessed. Noise texture shifted towards high frequency for Siemens Healthineers and Philips Healthcare but the opposite for GE Healthcare.