Figure 3:

The box plot of the RMSEs and SSIMs on the 100 testing slices. $\text{N}2{\text{N}}_{\text{d}}$ is the Noise2Noise denoising; $\text{N}2{\text{N}}_{\text{r}}$ is the Noise2Noise reconstruction; $\text{N}2{\text{N}}_{\text{r}}$(w/o pre) is the Noise2Noise reconstruction without pre-training with zero-image initialization; N2C is Noise2Clean (supervised learning). The SSIMs were calculated in the soft tissue window 40 ± 400 HU. Gaussian was used as baseline instead of FBP to eliminate the modeling bias between IR and analytical reconstruction. Compared to Noise2Noise reconstruction, Noise2Noise denoising does not fine-tune during testing.