Table 2. Performance comparison of the proposed UDAN with its six variations.
| Method | TRE (mm) ↓ | SSIM (%) ↑ | DICE (%) ↑ | TIME (s) ↓ | |Jϕ|≤ 0(%) ↓ |
|---|---|---|---|---|---|
| Baseline | 2.25 | 83.92 | 86.54 | 14 | 0.34 |
| Baseline with P-BLOCK | 2.21 | 84.27 | 87.15 | 17 | 0.29 |
| Baseline with C-BLOCK | 2.09 | 84.72 | 87.56 | 16 | 0.31 |
| Baseline with SP-BLOCK | 2.13 | 85.13 | 87.43 | 16 | 0.25 |
| Baseline with SC-BLOCK | 2.11 | 85.08 | 88.57 | 17 | 0.33 |
| UDAN | 2.07 | 86.34 | 89.74 | 18 | 0.28 |
The best results are shown bold font. 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; P-BLOCK, position attention block; SP-BLOCK, scale-aware position attention block; SC-BLOCK, scale-aware channel attention block; UDAN, unsupervised dual attention network.