Table 2.
Performance comparison of different methods on the Mindboggle dataset
| Method | DSC (%) | HD95 (mm) | ASSD (mm) |
 (%)
|
|---|---|---|---|---|
| Initial | 31.8 ± 2.4 | 7.12 ± 0.61 | 2.31 ± 0.25 | - |
| SPM | 49.2 ± 1.5 | 5.81 ± 0.43 | 1.58 ± 0.11 | - |
| SyN | 56.5 ± 1.4* | 5.69 ± 0.40 | 1.46 ± 0.10 | < 0.000006 |
| VoxelMorph | 53.6 ± 2.4* | 6.47 ± 0.56 | 1.69 ± 0.19 | < 0.2 |
| TransMorph | 58.6 ± 2.4* | 6.17 ± 0.55 | 1.53 ± 0.18 | < 0.9 |
| TransMatch | 57.2 ± 2.2* | 6.13 ± 0.55 | 1.54 ± 0.17 | < 0.9 |
| PR++ | 59.3 ± 2.0* | 6.11 ± 0.54 | 1.50 ± 0.17 | < 0.4 |
| PIVIT | 57.1 ± 1.5* | 5.66 ± 0.41 | 1.44 ± 0.11 | < 0.07 |
| Im2grid | 59.0 ± 1.6* | 5.72 ± 0.42 | 1.41 ± 0.12 | < 0.01 |
| NICETrans | 64.5 ± 1.3 | 5.39 ± 0.38 | 1.27 ± 0.10 | < 0.7 |
| ModeT | 61.9 ± 1.2* | 5.50 ± 0.35 | 1.32 ± 0.09 | < 0.02 |
| RDP | 64.5 ± 1.4* | 5.57 ± 0.43 | 1.30 ± 0.11 | < 0.006 |
| CorrMLP | 64.8 ± 1.2 | 5.40 ± 0.39 | 1.26 ± 0.10 | < 0.7 |
| CRR-Net | 64.8 ± 1.2 | 5.40 ± 0.39 | 1.26 ± 0.10 | < 0.3 |
Bold values indicate the best performance. *P < 0.05 compared with CRR-Net. DSC: Dice similarity coefficient; ASSD: Average symmetric surface distance; HD95: 95% Hausdorff distance; CRR-Net: Correlation reconstruction and refinement network