Table 3.
Mean upper limit of the noise level at which images were acceptable for clinical evaluation in non-denoised images and images denoised with DnCNN, SCNN and dDLR in five volunteers
| Reader A | Reader B | |||||||
|---|---|---|---|---|---|---|---|---|
| Noisy | DnCNN | SCNN | dDLR | Noisy | DnCNN | SCNN | dDLR | |
| T1WI | 3.8% (0.7) | 5.0% (0.6) | 5.4% (0.5) | 6.2% (0.4) | 3.2% (0.7) | 4.2% (0.7) | 4.8% (0.7) | 5.4% (0.5) |
| T2WI | 4.4% (0.5) | 6.8% (0.7) | 7.4% (0.8) | 8.6% (0.8) | 4.2% (0.7) | 6.6% (0.8) | 7.4% (0.8) | 8.6% (0.8) |
| FLAIR | 3.0% (0.6) | 4.0% (0.9) | 4.4% (0.5) | 4.8% (0.7) | 2.8% (0.7) | 3.6% (0.8) | 4.6% (0.5) | 4.8% (0.7) |
Data in parentheses are standard deviations. FLAIR, fluid-attenuated inversion recovery; DnCNN, denoising convolutional neural network; SCNN, shrinkage convolutional neural network; dDLR, deep learning-based reconstruction; T1WI, T1-weighted image; T2WI, T2-weighted image.