Table 1. Performance comparison of the proposed UDAN with six mainstream deformable image registration methods.
| Method | TRE (mm) ↓ | SSIM (%) ↑ | DICE (%) ↑ | TIME (s) ↓ | |Jϕ|≤ 0(%) ↓ |
|---|---|---|---|---|---|
| Affine | 4.27 | 68.17 | 64.38 | – | – |
| ANTs | 2.32 | 74.43 | 79.27 | 1123 | 0.16 |
| ELASTIX | 2.45 | 73.56 | 75.24 | 265 | 0.09 |
| B-spline | 3.27 | 69.49 | 70.49 | 943 | 0.06 |
| VTN | 2.12 | 85.48 | 88.61 | 16 | 0.12 |
| VoxelMorph | 2.17 | 85.34 | 87.94 | 14 | 0.34 |
| CycleMorph | 2.14 | 85.73 | 87.81 | 23 | 0.16 |
| UDAN | 2.07 | 86.34 | 89.74 | 18 | 0.28 |
The best results are shown bold font. The line of affine is the result of rigid registration before the application of deformable registration. UDAN, unsupervised dual attention network; TRE, target registration error; SSIM, structural similarity; DICE, dice similarity coefficient; |Jϕ|≤ 0(%), areas with negative Jacobian determinant are considered folding, the folded area as a percentage of the total area.