Table 11.
Performance of hybrid deep learning methods with and without the attention layers in their k-space deep learning components on the fastMRI dataset.
| Methods | NMSE (× 10−3) ↓ | PSNR ↑ | SSIM (× 10−2) ↑ | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Cartesian | Radial | Cartesian | Radial | Cartesian | Radial | |||||||
| mean±std | p-value | mean±std | p-value | mean±std | p-value | mean±std | p-value | mean±std | p-value | mean±std | p-value | |
| KIKI-Net | 18.02±0.02 | 7.5e-25 | 14.70± 0.02 | 3.8e-24 | 39.27±3.03 | 7.2e-36 | 40.51±3.30 | 6.8e-32 | 95.12±0.05 | 1.6e-22 | 95.29±0.05 | 4.4e-19 |
| w/attention | 17.84±0.01 | - | 14.38±0.01 | - | 39.36±2.53 | - | 40.65±3.06 | - | 95.18±0.05 | - | 95.35±0.04 | - |
| Hybrid-Net | 17.89±0.01 | 6.3e-22 | 14.66±0.01 | 1.8e-19 | 39.32±2.63 | 4.9e-27 | 40.59±3.04 | 9.2e-27 | 95.15±0.05 | 1.4e-22 | 95.32±0.05 | 2.8e-16 |
| w/attention | 17.64±0.01 | 14.38±0.01 | - | 39.40±2.74 | - | 40.69±2.86 | - | 95.20±0.04 | - | 95.39±0.04 | - | |