Real Pairs Are Not Enough: Physically Anchored Supervision for Volumetric CT Super-Resolution
Organizations: Chengdu University of Technology · The Chinese University of Hong Kong · Specreation Instruments Co., Ltd.
Abstract
Volumetric CT super-resolution (SR) models are commonly trained on synthetic or physically acquired pairs of low-resolution (LR) and high-resolution (HR) volumes. However, synthetic pairs miss real acquisition effects, whereas registered physical pairs retain inter-scan noise, artifact, and microstructure differences---causing direct voxelwise regression, and even naive synthetic-real mixing, to oversmooth fine porous structures. We introduce the Physically-anchored Implicit Network (PIN) with Real-Pair Dual-Anchor learning, which decouples physical-pair supervision into two complementary roles. The HR anchor teaches fine structure from registered HR through pair-conditioned spectral replay, while the LR anchor constrains the prediction from acquired LR via affine-calibrated low-pass projection. Together with exact synthetic supervision in one shared volumetric implicit model, PIN learns from both physical scans without forcing the acquired-LR reconstruction to match registered HR voxel by voxel. Inference requires only acquired LR. Across 30 external baseline configurations on real CT, PIN achieves the lowest full-volume LPIPS, reduces average symmetric surface distance by 42.3-49.5% and bone-volume-fraction error by 52.2-53.4% relative to the three strongest baselines for each metric, and lowers the erased reference-pore volume fraction from 37.3-72.9% to 10.4%. Code, pretrained models, and evaluation scripts will be publicly available.
Figures & tables
| Baseline | Real | Common obj./ROI | ASD HR SR PIN / baseline | ASD SR HR PIN / baseline | ASSD PIN / baseline | HD95 PIN / baseline |
|---|---|---|---|---|---|---|
| MetaSR | ✓ | 10/32 | 1.187 / 5.632 | 0.544 / 0.428 | 0.965 / 5.114 | 3.280 / 13.721 |
| MetaSR | ✗ | 10/33 | 1.185 / 4.642 | 0.542 / 0.544 | 0.963 / 4.084 | 3.276 / 12.042 |
| LIIF | ✓ | 10/31 | 1.164 / 5.856 | 0.545 / 0.400 | 0.943 / 5.288 | 3.248 / 14.297 |
| LIIF | ✗ | 10/33 | 1.185 / 4.654 | 0.542 / 0.545 | 0.963 / 4.095 | 3.276 / 12.100 |
| LTE | ✓ | 10/30 | 1.167 / 5.697 | 0.547 / 0.373 | 0.946 / 5.107 | 3.255 / 14.102 |
| LTE | ✗ | 10/33 | 1.185 / 4.572 | 0.542 / 0.543 | 0.963 / 4.019 | 3.276 / 11.873 |
| Configuration | Real | ASD HR SR | ASD SR HR | ASSD | HD95 | Finite objects | Finite ROIs |
|---|---|---|---|---|---|---|---|
| MetaSR | ✓ | 5.632 (32/40 ROI) | 0.428 (32/40 ROI) | 5.114 (32/40 ROI) | 13.721 (32/40 ROI) | 10/10 | 32/40 |
| MetaSR | ✗ | 4.642 (33/40 ROI) | 0.544 (33/40 ROI) | 4.084 (33/40 ROI) | 12.042 (33/40 ROI) | 10/10 | 33/40 |
| LIIF | ✓ | 5.856 (31/40 ROI) | 0.400 (31/40 ROI) | 5.288 (31/40 ROI) | 14.297 (31/40 ROI) | 10/10 | 31/40 |
| LIIF | ✗ | 4.654 (33/40 ROI) | 0.545 (33/40 ROI) | 4.095 (33/40 ROI) | 12.100 (33/40 ROI) | 10/10 | 33/40 |
| LTE | ✓ | 5.697 | 0.373 | 5.107 | 14.102 | 10/10 | 30/40 |
| LTE | ✗ | 4.572 (33/40 ROI) | 0.543 (33/40 ROI) | 4.019 (33/40 ROI) | 11.873 (33/40 ROI) | 10/10 | 33/40 |
| Configuration | Real | Finite ROIs | Boundary QC | Insufficient halo | Empty SR | Other invalid |
|---|---|---|---|---|---|---|
| MetaSR | ✓ | 32/40 | 8 | 3 | 0 | 0 |
| MetaSR | ✗ | 33/40 | 7 | 0 | 0 | 0 |
| LIIF | ✓ | 31/40 | 9 | 4 | 0 | 0 |
| LIIF | ✗ | 33/40 | 7 | 0 | 0 | 0 |
| LTE | ✓ | 30/40 | 10 | 4 | 0 | 0 |
| LTE | ✗ | 33/40 | 7 | 0 | 0 | 0 |
| Configuration | Real | Volume PSNR | Volume SSIM | Volume LPIPS | ROI PSNR | ROI SSIM | ROI LPIPS |
|---|---|---|---|---|---|---|---|
| MetaSR | ✓ | 16.038 | 0.3976 | 0.4996 | 18.536 | 0.5451 | 0.3829 |
| MetaSR | ✗ | 15.295 | 0.3383 | 0.5406 | 17.223 | 0.4753 | 0.4206 |
| LIIF | ✓ | 16.011 | 0.3849 | 0.4954 | 18.411 | 0.5210 | 0.3813 |
| LIIF | ✗ | 15.283 | 0.3377 | 0.5432 | 17.165 | 0.4745 | 0.4211 |
| LTE | ✓ | 16.028 | 0.3977 | 0.4856 | 18.461 | 0.5407 | 0.3691 |
| LTE | ✗ | 15.281 | 0.3384 | 0.5455 | 17.165 | 0.4751 | 0.4204 |
| Configuration | Real | SDice 0.5 | SDice 1 | SDice 2 | HRcov 0.5 | HRcov 1 | HRcov 2 |
|---|---|---|---|---|---|---|---|
| MetaSR | ✓ | 0.2423 | 0.4213 | 0.5011 | 0.1834 | 0.3486 | 0.4368 |
| MetaSR | ✗ | 0.1981 | 0.3772 | 0.4912 | 0.1334 | 0.2877 | 0.4067 |
| LIIF | ✓ | 0.2324 | 0.3980 | 0.4709 | 0.1751 | 0.3288 | 0.4084 |
| LIIF | ✗ | 0.1989 | 0.3777 | 0.4938 | 0.1340 | 0.2879 | 0.4090 |
| LTE | ✓ | 0.2378 | 0.4006 | 0.4744 | 0.1783 | 0.3293 | 0.4103 |
| LTE | ✗ | 0.2012 | 0.3815 | 0.4996 | 0.1357 | 0.2912 | 0.4149 |
| Configuration | Real | Dice | Jaccard | Precision | Recall | SRsup 0.5 | SRsup 1 | SRsup 2 |
|---|---|---|---|---|---|---|---|---|
| MetaSR | ✓ | 0.8750 | 0.7792 | 0.8208 | 0.9424 | 0.6145 | 0.9319 | 0.9908 |
| MetaSR | ✗ | 0.8491 | 0.7403 | 0.8134 | 0.8928 | 0.5526 | 0.8846 | 0.9809 |
| LIIF | ✓ | 0.8762 | 0.7811 | 0.8204 | 0.9466 | 0.6415 | 0.9345 | 0.9889 |
| LIIF | ✗ | 0.8495 | 0.7410 | 0.8131 | 0.8941 | 0.5523 | 0.8832 | 0.9810 |
| LTE | ✓ | 0.8775 | 0.7832 | 0.8194 | 0.9506 | 0.6539 | 0.9370 | 0.9909 |
| LTE | ✗ | 0.8494 | 0.7408 | 0.8133 | 0.8936 | 0.5531 | 0.8839 | 0.9816 |
| Configuration | Real | Unsupported 0.5 | Unsupported 1 | Unsupported 2 | Missing 0.5 | Missing 1 | Missing 2 |
|---|---|---|---|---|---|---|---|
| MetaSR | ✓ | 17.92 | 2.68 | 0.12 | 5.76 | 1.69 | 0.61 |
| MetaSR | ✗ | 18.66 | 3.19 | 0.23 | 10.72 | 6.68 | 3.88 |
| LIIF | ✓ | 17.96 | 2.71 | 0.16 | 5.34 | 1.58 | 0.60 |
| LIIF | ✗ | 18.69 | 3.21 | 0.23 | 10.59 | 6.54 | 3.73 |
| LTE | ✓ | 18.06 | 2.73 | 0.12 | 4.94 | 1.43 | 0.56 |
| LTE | ✗ | 18.67 | 3.21 | 0.23 | 10.64 | 6.55 | 3.69 |
| Configuration | Checkpoint | ASSD ( ) | HD95 ( ) | Surface Dice @1 | HR-covered @1 | Finite ROIs |
|---|---|---|---|---|---|---|
| Native support: bounded scores on 40/40 ROIs | ||||||
| mDCSRN | Released | 1.724 | 5.855 | 0.5669 | 0.4568 | 38/40 |
| mDCSRN | Direct retrained | 1.750 | 5.932 | 0.5662 | 0.4567 | 39/40 |
| MTVNet | Released | 1.839 | 6.183 | 0.5537 | 0.4453 | 38/40 |
| MTVNet | Direct retrained | 1.753 | 5.877 | 0.5628 | 0.4564 | 38/40 |
| RRDBNet3D | Released | 1.968 | 6.538 | 0.5393 | 0.4308 | 38/40 |
| SR architecture | Type | Parameters (M) | Source |
|---|---|---|---|
| PIN | 3D | 15.499361 | Counted from full SR network |
| MetaSR | 2D | 1.665560 | Counted from full SR network |
| LIIF | 2D | 1.567323 | Counted from full SR network |
| LTE | 2D | 1.714267 | Counted from full SR network |
| SRNO | 2D | 2.015323 | Counted from full SR network |
| CiaoSR | 2D | 2.685982 | Counted from full SR network |
| Metric | Method | Seed 0 | Seed 1 | Seed 2 | Mean | SD |
|---|---|---|---|---|---|---|
| PSNR-2D (dB) | PIN | 15.09032 | 15.14337 | 15.17161 | 15.13510 | 0.04127 |
| SSIM-2D | PIN | 0.38776 | 0.38593 | 0.38995 | 0.38788 | 0.00201 |
| LPIPS | PIN | 0.45140 | 0.44769 | 0.44830 | 0.44913 | 0.00199 |
| ASSD ( ) | PIN | 0.99465 | 0.98236 | 1.00539 | 0.99414 | 0.01152 |
| HD95 ( ) | PIN | 3.47293 | 3.39150 | 3.46872 | 3.44438 | 0.04585 |
| Surface Dice @1 | PIN | 0.70460 | 0.70635 | 0.70179 | 0.70425 | 0.00230 |
| Metric | Method | Seed 0 | Seed 1 | Seed 2 | Mean | SD |
|---|---|---|---|---|---|---|
| ASD HR to SR ( ) | PIN | 1.23056 | 1.21325 | 1.24438 | 1.22940 | 0.01560 |
| ASD SR to HR ( ) | PIN | 0.55568 | 0.55652 | 0.54682 | 0.55301 | 0.00538 |
| Surface Dice @0.5 | PIN | 0.40207 | 0.40209 | 0.40007 | 0.40141 | 0.00116 |
| Surface Dice @1.0 | PIN | 0.70460 | 0.70635 | 0.70179 | 0.70425 | 0.00230 |
| Surface Dice @2.0 | PIN | 0.86866 | 0.86992 | 0.86291 | 0.86717 | 0.00373 |
| HR-covered @0.5 | PIN | 0.32824 | 0.32804 | 0.32494 | 0.32708 | 0.00185 |
| ( ) | PIN | Best external | Real | External | Margin (pp) | |
|---|---|---|---|---|---|---|
| -0.050 | 0.5 | 0.3811 | mDCSRN | ✗ | 0.2890 | +9.21 |
| -0.050 | 1 | 0.6762 | mDCSRN | ✗ | 0.5017 | +17.45 |
| -0.050 | 2 | 0.8416 | mDCSRN | ✗ | 0.6240 | +21.76 |
| -0.025 | 0.5 | 0.3930 | mDCSRN | ✗ | 0.3096 | +8.34 |
| -0.025 | 1 | 0.6930 | mDCSRN | ✗ | 0.5392 | +15.38 |
| -0.025 | 2 | 0.8579 | mDCSRN | ✗ | 0.6703 | +18.76 |
| ( ) | PIN | Best external | Real | External | Margin (pp) | |
|---|---|---|---|---|---|---|
| -0.050 | 0.5 | 0.3175 | mDCSRN | ✗ | 0.2127 | +10.48 |
| -0.050 | 1 | 0.5926 | mDCSRN | ✗ | 0.3954 | +19.72 |
| -0.050 | 2 | 0.7782 | mDCSRN | ✗ | 0.5283 | +24.99 |
| -0.025 | 0.5 | 0.3241 | mDCSRN | ✗ | 0.2291 | +9.50 |
| -0.025 | 1 | 0.6056 | mDCSRN | ✗ | 0.4308 | +17.48 |
| -0.025 | 2 | 0.7978 | mDCSRN | ✗ | 0.5791 | +21.87 |
| Configuration | Real | Tb.Th applicable | Tb.Th HR finite | Tb.Th SR finite | Solid applicable | Solid HR finite | Solid SR finite |
|---|---|---|---|---|---|---|---|
| MetaSR | ✓ | 2 | 2 | 2/2 | 8 | 8 | 2/8 |
| MetaSR | ✗ | 2 | 2 | 2/2 | 8 | 8 | 2/8 |
| LIIF | ✓ | 2 | 2 | 2/2 | 8 | 8 | 3/8 |
| LIIF | ✗ | 2 | 2 | 2/2 | 8 | 8 | 3/8 |
| LTE | ✓ | 2 | 2 | 2/2 | 8 | 8 | 3/8 |
| LTE | ✗ | 2 | 2 | 2/2 | 8 | 8 | 3/8 |
| Configuration | Real | FP/ (%) | FN/ (%) | Signed solid-volume bias (pp) | Absolute VF error (dimensionless) |
|---|---|---|---|---|---|
| Supervision strategies and component controls | |||||
| Synth-only | – | 8.9018 | 3.2577 | 5.6441 | 0.0636 |
| Real-only | – | 8.7519 | 1.9514 | 6.8005 | 0.0707 |
| Direct Mix | – | 8.8074 | 1.7685 | 7.0389 | 0.0720 |
| PIN | – | 5.5857 | 9.2413 | -3.6556 | 0.0411 |
| A0 | – | 5.6194 | 9.2180 | -3.5987 | 0.0409 |
| Method | Real | Fully erased reference-pore volume share | Lost HR face paths |
|---|---|---|---|
| Shared-backbone supervision strategies | |||
| Synth-only | – | 68.89% | 108/250 |
| Real-only | – | 53.90% | 64/250 |
| Direct Mix | – | 53.43% | 65/250 |
| PIN | – | 10.41% | 39/250 |
| Complete external benchmark: 15 families, Synthetic and Real | |||
| Source | Perturbation | Geometry | Copy | Structural | Energy | Active fraction | |
|---|---|---|---|---|---|---|---|
| A0 | Healthy control | 0.0000 | 0.0000 | 0.0085 | 1.9026 | 0.0000 | 0.2000 |
| A0 | Gaussian blur | 0.0279 | 0.0003 | 0.0361 | 1.8519 | 0.0068 | 0.4500 |
| A0 | Gaussian noise | 0.0029 | 0.0000 | 0.0078 | 16.3889 | 0.0000 | 0.2781 |
| A0 | Grid, period 4 | 0.0046 | 0.0000 | 0.0095 | 3.4227 | 0.0000 | 0.2281 |
| A0 | Grid, period 8 | -0.0000 | 0.0000 | 0.0086 | 1.6210 | 0.0000 | 0.1938 |
| A0 | Grid, period 16 | 0.0008 | 0.0000 | 0.0087 | 2.2992 | 0.0000 | 0.2375 |
| Source | Perturbation | Cap-hit fraction | |
|---|---|---|---|
| A0 | Healthy control | 0.0476 | 0.0375 |
| A0 | Gaussian blur | 0.0457 | 0.0750 |
| A0 | Gaussian noise | 0.1000 | 0.2500 |
| A0 | Grid, period 4 | 0.0721 | 0.2875 |
| A0 | Grid, period 8 | 0.0412 | 0.0125 |
| A0 | Grid, period 16 | 0.0559 | 0.1250 |
Medical Image Super-Resolution
Medical image super-resolution reconstructs fine spatial detail from low-resolution scans to support clearer anatomy and more useful clinical measurements.
See all 62 papers