. Author manuscript; available in PMC: 2021 Nov 12.

Published in final edited form as: IEEE Trans Med Imaging. 2021 Oct 27;40(11):3154–3164. doi: 10.1109/TMI.2021.3076191

TABLE III.

Comparison of Denoising Performance on Different Motion-Compensated Images. Our MDPET Is Compared With 1) SAN and 2) Two-Stage Processing Methods That Consist of Motion Estimation and Denoising (DN).

Method-mean(std)	NMAE	SSIM	PSNR
✗REG+✗DN	.1712(.0225)	.9018(.0175)	25.87(1.87)
✓NRB+✗DN	.1174(.0198)	.9424(.0096)	28.97(1.79)
NRB+UNet	.1166(.0177)	.9479(.0068)	29.49(1.85)
NRB+GAN	.1147(.0179)	.9489(.0071)	29.66(1.91)
✓VM+✗DN	.1165(.0130)	.9431(.0080)	28.98(1.90)
VM+UNet	.1125(.0124)	.9480(.0052)	29.43(1.99)
VM+GAN	.1128(.0130)	.9490(.0061)	29.48(1.98)
✓SAN+✗DN	.1401(.0187)	.9191(.0154)	27.99(1.49)
SAN+UNet	.1062(.0122)	.9498(.0061)	30.31(1.87)
SAN+GAN	.1036(.0117)	.9503(.0061)	30.87(1.79)
✓Ours+✗DN	.1383(.0185)	.9193(.0153)	28.14(1.46)
Ours	.0883(.0133)	.9669(.0054)	32.28(1.89)

✓ and ✗ Denote Use or Not Use of a Specific Processing Stage. For Example, ✓NRB+✗DN Means NRB Is Used For Estimating the Motion and Generating The Averaged Image, but No Denoising Step Is Applied. The Corresponding Boxplot Comparison Results With Statistical Analysis Are Shown in Figure 7