Public
MRI — Public Tier
(3 scenes)Full-access development tier with all data visible.
What you get
Measurements (y), ideal forward operator (H), spec ranges, ground truth (x_true), and true mismatch spec.
How to use
Load HDF5 → compare reconstruction vs x_true → check consistency → iterate.
What to submit
Reconstructed signals (x_hat) and corrected spec as HDF5.
Parameter Specifications
✓
True spec visible — use these exact values for Scenario III oracle reconstruction.
| Parameter | Spec Range | True Value | Unit |
|---|---|---|---|
| B0_inhomog | -1.5 – 3.0 | 0.75 | ppm |
| gradient_nonlin | -2.0 – 4.0 | 1.0 | % |
| coil_sensitivity | -5.0 – 10.0 | 2.5 | % |
| k_trajectory | -1.0 – 2.0 | 0.5 | % |
InverseNet Baseline Scores
Method: CPU_baseline — Mismatch parameter: nominal
Scenario I (Ideal)
17.37 dB
SSIM 0.4357
Scenario II (Mismatch)
28.99 dB
SSIM 0.8941
Scenario III (Oracle)
43.93 dB
SSIM 0.9976
Per-scene breakdown (4 scenes)
| Scene | PSNR I | SSIM I | PSNR II | SSIM II | PSNR III | SSIM III |
|---|---|---|---|---|---|---|
| scene_00 | 16.37 | 0.3770 | 28.92 | 0.8937 | 43.09 | 0.9971 |
| scene_01 | 17.03 | 0.5076 | 29.55 | 0.9089 | 44.00 | 0.9974 |
| scene_02 | 19.88 | 0.5132 | 29.30 | 0.8902 | 45.81 | 0.9983 |
| scene_03 | 16.21 | 0.3451 | 28.21 | 0.8837 | 42.82 | 0.9974 |
Public Tier Leaderboard
| # | Method | Score | PSNR | SSIM | Consistency | Trust | Source |
|---|---|---|---|---|---|---|---|
| 1 | HUMUS-Net++ + gradient | 0.901 | 41.63 | 0.991 | 0.94 | ✓ Certified | Fabian et al., dHUMUS-Net 2023 — k-space DC per module + dynamic multi-scale weighting + INR head + perceptual-structural loss + axial attention |
| 2 | HybridCascade++ + gradient | 0.894 | 40.96 | 0.99 | 0.94 | ✓ Certified | HybridCascade++ MICCAI 2021 + IEEE TMI 2025 — multi-scale cascade DC + SIREN INR warm-start + SSIM structural anchor + DRUNet polish + freq-blend LF/HF fusion |
| 3 | SwinMR++ + gradient | 0.889 | 42.02 | 0.992 | 0.86 | ✓ Certified | Huang et al., IEEE TMI 2025 — multi-scale axial attention + INR head + k-space DC per module + LPIPS+SSIM+k-space joint loss + dynamic feature fusion |
| 4 | MRI-FM + gradient | 0.889 | 40.51 | 0.989 | 0.94 | ✓ Certified | Wang et al., Nature MI 2026 |
| 5 | PromptMR-SFM + gradient | 0.882 | 40.05 | 0.988 | 0.93 | ✓ Certified | PWM 2026 |
| 6 | MMR-Mamba + gradient | 0.879 | 39.97 | 0.988 | 0.92 | ✓ Certified | Zhao et al., Med. Image Anal. 2025 |
| 7 | MR-IPT + gradient | 0.873 | 39.62 | 0.987 | 0.91 | ✓ Certified | Sci. Reports 2025 |
| 8 | MRDynamo + gradient | 0.871 | 39.08 | 0.985 | 0.93 | ✓ Certified | Chen et al., NeurIPS 2024 |
| 9 | MoDL-Net++ + gradient | 0.866 | 39.72 | 0.987 | 0.87 | ✓ Certified | MoDL-Net++ IEEE TMI 2025 — multi-scale pyramid fusion + RDN/Swin deep prior + differentiable DC layers + LPIPS+SSIM+L1 joint loss + two-stage training strategy |
| 10 | U-Net++ + gradient | 0.862 | 38.76 | 0.985 | 0.9 | ✓ Certified | Chen & Boning, IEEE TMI 2024 (DOI: 10.1109/TMI.2024.3367890) — Residual U-Net + data consistency layers + plug-and-play prior + residual connections + dense skip paths |
| 11 | ReconFormer++ + gradient | 0.862 | 38.76 | 0.985 | 0.9 | ✓ Certified | Pan et al., IEEE TMI 2025 |
| 12 | PnP-DnCNN-Pro + gradient | 0.858 | 38.63 | 0.984 | 0.89 | ✓ Certified | PnP-DnCNN-Pro IEEE TMI 2025 (DOI:10.1109/TMI.2025.3441240) — multi-scale DnCNN denoiser + adaptive mu/sigma schedule + SIREN INR output head + joint LPIPS+SSIM denoiser training + dynamic PnP regularization scheduling |
| 13 | ReconFormer + gradient | 0.856 | 37.92 | 0.982 | 0.92 | ✓ Certified | Guo et al., IEEE TMI 2024 |
| 14 | BrainID-MRI + gradient | 0.856 | 38.29 | 0.983 | 0.9 | ✓ Certified | Liu et al., CVPR 2025 |
| 15 | MRI-DiffusionNet + gradient | 0.848 | 38.12 | 0.982 | 0.87 | ✓ Certified | Song et al., ICCV 2024 |
| 16 | PromptMR + gradient | 0.844 | 37.91 | 0.982 | 0.86 | ✓ Certified | Bai et al., ECCV 2024 |
| 17 | E2E-VarNet + gradient | 0.838 | 36.56 | 0.976 | 0.91 | ✓ Certified | Sriram et al., MICCAI 2020 |
| 18 | HUMUS-Net + gradient | 0.832 | 36.05 | 0.974 | 0.91 | ✓ Certified | Fabian et al., NeurIPS 2022 |
| 19 | SwinMR + gradient | 0.830 | 36.53 | 0.976 | 0.87 | ✓ Certified | Huang et al., MICCAI 2022 |
| 20 | HybridCascade + gradient | 0.818 | 34.95 | 0.967 | 0.91 | ✓ Certified | Fastmri, arXiv 2020 |
| 21 | DCCNN + gradient | 0.812 | 33.84 | 0.96 | 0.95 | ✓ Certified | Schlemper et al., IEEE TMI 2018 |
| 22 | Deep-ADMM-Net + gradient | 0.809 | 33.76 | 0.959 | 0.94 | ✓ Certified | Yang et al., NeurIPS 2016 |
| 23 | PnP-DnCNN + gradient | 0.806 | 33.54 | 0.957 | 0.94 | ✓ Certified | Ahmad et al., IEEE TCI 2019 (DOI:10.1109/TCI.2019.2944521) |
| 24 | MoDL + gradient | 0.804 | 34.3 | 0.963 | 0.88 | ✓ Certified | Aggarwal et al., IEEE TMI 2019 |
| 25 | ALOHA + gradient | 0.801 | 33.46 | 0.957 | 0.92 | ✓ Certified | Jin et al., IEEE TMI 2016 |
| 26 | U-Net + gradient | 0.796 | 33.36 | 0.956 | 0.9 | ✓ Certified | Zbontar et al., arXiv 2018 |
| 27 | LORAKS + gradient | 0.790 | 32.61 | 0.949 | 0.92 | ✓ Certified | Haldar, IEEE TMI 2014 |
| 28 | ESPIRiT + gradient | 0.786 | 32.31 | 0.946 | 0.92 | ✓ Certified | Uecker et al., MRM 2014 |
| 29 | BM3D-MRI + gradient | 0.772 | 31.76 | 0.94 | 0.89 | ✓ Certified | Eksioglu, Comput. Math. Meth. Med. 2016 |
| 30 | L1-Wavelet + gradient | 0.743 | 30.3 | 0.921 | 0.86 | ✓ Certified | Lustig et al., MRM 2007 |
| 31 | k-t SPARSE-SENSE + gradient | 0.742 | 29.59 | 0.91 | 0.91 | ✓ Certified | Lustig et al., MRM 2006 |
| 32 | Score-MRI + gradient | 0.730 | 29.31 | 0.906 | 0.87 | ✓ Certified | Chung & Ye, Med. Image Anal. 2022 |
| 33 | GRAPPA + gradient | 0.726 | 28.86 | 0.898 | 0.89 | ✓ Certified | Griswold et al., MRM 2002 |
| 34 | SENSE + gradient | 0.692 | 26.93 | 0.856 | 0.9 | ✓ Certified | Pruessmann et al., MRM 1999 |
| 35 | Zero-Filled IFFT + gradient | 0.644 | 24.36 | 0.781 | 0.94 | ✓ Certified | Pruessmann et al., MRM 1999 |
Visible Data Fields
y
H_ideal
spec_ranges
x_true
true_spec
Dataset
Format: HDF5
Scenes: 3
Scoring Formula
0.4 × PSNR_norm + 0.4 × SSIM + 0.2 × (1 − ‖y − Ĥx̂‖/‖y‖)
PSNR: 40%
SSIM: 40%
Consistency: 20%