. Author manuscript; available in PMC: 2021 Feb 26.

Published in final edited form as: IEEE Trans Comput Imaging. 2020 Apr 27;6:843–856. doi: 10.1109/TCI.2020.2990299

Table I. Quantitative measures and computational resources for CBCT reconstructions of the walnut data from 60 angles. Computed values are given for the test data in comparison to the ground-truth from full measurements. Additionally, we present estimated benchmark values for LGS.

	PSNR	SSIM	Train	Exec.	Parameter	Memory
FDK	26.95	0.424	~1m	260ms	1	723MB
U-Net denoiser	34.62	0.910	4h29m	528ms	6.3·10⁶	6097MB
MS-LFGS (ResNet)	32.98	0.878	5h04m	526ms	2.4·10⁴	2547MB
MS-LFGS (MINI U-Net)	34.13	0.903	7h02m	645ms	2.1·10⁵	4853MB
∂ U-Net	34.69	0.914	7h28m	795ms	4.1·10⁶	5313MB
LGS (5 ITER., ESTIMATED)	–	–	~25h	2.5s	~ 2 · 10⁵	~ 12500MB