Calibrating the Scanner, Not the Prior: Per-Scan Forward-Operator Estimation for Cross-Vendor MRI Reconstruction

Deep networks that reconstruct accelerated MRI scans from raw scanner measurements work well on the scanner they were trained on and often break on a different vendor's scanner. The 2020 fastMRI challenge traced part of this "transfer wall" to something mundane. GE scanners use different default readout oversampling than Siemens scanners, so models that had implicitly learned Siemens k-space dimensions failed on GE data (Results of the 2020 fastMRI Challenge). K-space is the raw frequency-domain data an MRI scanner records. One family of methods is supposed to be immune to this. It keeps a learned model of what clean anatomy looks like (the "prior") separate from an explicit mathematical model of how the scanner measures it (the "forward operator"). Those methods, however, assume the forward operator is known exactly. Vendor shift is precisely the case where it is not.

We propose to freeze the learned anatomy prior and instead estimate, for each individual scan, a small set of physically meaningful operator parameters. These are the parameters that differ across vendors: readout field of view and oversampling, vendor k-space filtering, coil-sensitivity errors, and phase offsets. They are estimated jointly with the image during reconstruction, and their uncertainty is carried into the output. The claim is testable and falsifiable. If cross-vendor failure is mostly operator mis-specification, then adapting a handful of operator parameters should recover most of what adapting millions of prior weights recovers. It should do so faster, with no change to certified weights, and with an explicit signal of when the operator cannot be pinned down.

Core Claim

Existing decoupled methods put their adaptation in the wrong place. AdaDiff fine-tunes the diffusion prior's weights on each test subject's k-space. A diffusion model here means a generator that learns to turn noise into images by repeated denoising. AdaDiff reports gains of 7.7 dB PSNR and 20.3% SSIM over traditional methods under distribution shift. FedGIMP optimizes a GAN's latent code and noise variables against a data-consistency loss $\|y - A(G(z))\|_2^2$. It reports +6.93 dB PSNR over conditional federated baselines under multi-coil configuration shift. Both treat $A$ as fixed and correct. When $A$ is wrong, both methods can reach low data-consistency loss by bending the image, or the prior itself, to explain what is really operator error. That is the mechanism by which a flexible prior produces plausible-looking wrong structure.

Our claim is that a per-scan operator model with a few dozen parameters is enough to capture the dominant vendor-specific physics. With such a model, a locked prior can match or beat weight adaptation on cross-vendor transfer. The disagreement among plausible operator estimates is also a useful warning signal for hallucination.

The closest methodological ancestor comes from segmentation rather than reconstruction. SAMSEG for multiple sclerosis keeps an anatomical shape prior fixed and fits scanner-specific intensity parameters per scan by expectation-maximization. It reaches lesion Dice of 0.57–0.65 across Siemens and Philips cohorts without retraining. We carry that "fixed anatomy, per-scan acquisition parameters" logic down to the raw-data level. There the acquisition parameters are operator physics rather than Gaussian intensity mixtures.

Research Program

The forward model we fit is

$$y = M\,W_\phi\,\mathcal{F}_\phi\,S_\phi\,x + \epsilon.$$

In this model, $x$ is the image, $S_\phi$ applies coil sensitivities (an ESPIRiT-style initial estimate times a low-order smooth multiplicative correction per coil), and $\mathcal{F}_\phi$ is the Fourier transform with a readout field-of-view and oversampling parameter. $W_\phi$ is a vendor k-space weighting (apodization), expanded in a few smooth radial basis functions. $M$ is the known sampling mask, and $\epsilon$ is noise. The vector $\phi$ also includes a global phase and a linear phase ramp, and it totals roughly 20–60 numbers per scan. The prior is a diffusion model trained once, on Siemens fastMRI data only, and then locked.

Reconstruction follows diffusion posterior sampling, the approach of combining an unconditional diffusion prior with a measurement model at each denoising step. CT reconstruction with a nonlinear measurement model uses it with nonlinear CT physics. In our version, each data-consistency step on $x$ is interleaved with a gradient step on $\phi$ under a weak prior over physically admissible ranges. Several independent chains are run. The spread of $\phi$ across chains gives an operator-uncertainty estimate, and the pixelwise spread of $x$ is decomposed into a part explained by operator variation and a residual part. This design is not tied to Gaussian-noise diffusion. The same outer loop can wrap the Fourier-constrained diffusion bridge, which replaces Gaussian noise with structured frequency removal, and we test both prior types.

The main technical risk is identifiability. A smooth k-space filter $W_\phi$ trades off against image blur, and a smooth coil correction trades off against image intensity inhomogeneity. We rely on two asymmetries. First, a prior trained on sharp, fully sampled images assigns low probability to blurred ones, so blur should be pushed into $W_\phi$. Second, coil corrections are shared across slices of a volume, while anatomy is not. We will check identifiability directly with the injected-shift experiments below, where the true $\phi$ is known.

Reverse Engineering Breast MRIs shows that vendor and scanner model are recoverable from pixels with 99.74% and 97.78% accuracy. We use a small classifier of this kind only to initialize $\phi$ from a vendor-level default. Removing it is an ablation, since DICOM metadata is often stripped and our method must not depend on it.

Baselines

The naive reference is a static diffusion posterior sampler with the nominal Siemens operator and no adaptation of anything. AdaDiff shows this degrades under shift, and our ablation reuses exactly this configuration with $\phi$ frozen at nominal values. The difference between the two isolates the contribution of operator estimation. At the other end, an oracle run with $\phi$ set to the true injected values (in simulation) or to vendor-provided parameters (on real data) gives the ceiling for operator correction.

The strongest competitor is AdaDiff's per-subject prior-weight adaptation. It is the only method in the file that directly targets subject-level shift with a decoupled prior, and it reports 7.7 dB PSNR and 20.3% SSIM gains over traditional methods when the image distribution shifts. We re-implement it on the same Siemens-trained prior, with the same number of prior network evaluations per scan and the same hyperparameter-search budget on a held-out Siemens validation split. We also run it combined with our operator estimation. Beating it means matching its cross-vendor PSNR/SSIM within 0.3 dB while using no weight updates. Alternatively, it means exceeding it on scans where we inject operator errors that weight adaptation can only absorb by distorting anatomy.

The second competitor is FedGIMP-style latent optimization with a fixed operator (FedGIMP). It represents adapting in latent space rather than weights. We use a GAN trained on the same data and give it a matched iteration budget.

The third competitor is an unrolled network with data-consistency layers of the variational-network type used by top fastMRI 2020 entries. The top 4× entry reached SSIM around 0.964 in-domain (fastMRI 2020). The file gives no numeric transfer-track SSIM, only that it was "significantly lower"; no reported result found on Lacuna for transfer-track SSIM. For context in the high-acceleration regime, MDPG reports PSNR 30.43 dB and SSIM 0.740 on fastMRI knee at 8×, versus 29.75 dB for ISTA-Net, but it evaluates in-domain only.

Why not just…?

Why not just fix the oversampling in preprocessing? This is the most serious threat. If the transfer wall is mostly a readout-oversampling convention, a deterministic pipeline could remove it. That pipeline would crop the readout to a common field of view using header metadata, re-estimate coil maps with ESPIRiT, and normalize intensity. We therefore include "metadata harmonization plus static sampler" as a full baseline, not a strawman. If it closes over 80% of the gap between the naive sampler and the oracle on real cross-vendor data, we will narrow our claim to settings with missing or misreported metadata and to residual effects such as filters and phase. Those residual settings are common after anonymization, as the reverse-engineering paper notes.

Why not just adapt the prior weights, as AdaDiff does? Weight adaptation is flexible enough to absorb operator error, and that is exactly the problem. It can reach data consistency by moving anatomy. It also costs a backward pass through a large network per subject and breaks the locked-weight requirement that INPUTADAPTER documents for certified software. In clinical time series, INPUTADAPTER shows that adapting inputs to a frozen predictor can match fine-tuning. Our injected-shift experiment settles the image-quality question: if AdaDiff matches us on image metrics but its error concentrates in structured anatomy while ours does not, the case for adapting the operator holds.

Why not train an unrolled network with heavy augmentation over operator variations? Physics-augmented supervised training might learn invariance to filters and oversampling directly. We will train such a network with $\phi$ randomized over the same ranges our method searches. This comparison is fair and could win on speed. Our expected advantage lies in shifts outside the augmentation range and in the uncertainty signal, which a single forward network does not provide. OptTTA is the analogous story in segmentation: it gained 22.5% Dice on M&Ms by searching transformations at test time rather than trusting training augmentation. Its transformations, however, are heuristic image operations, whereas ours are physical.

Proposed Experiments

The first experiment is controlled operator injection on multi-coil fastMRI brain and knee data (fastMRI). We apply known readout crops, apodization filters, coil-map perturbations, and phase ramps at 4× and 8× acceleration. We measure PSNR, SSIM, NMSE, and error in recovering $\phi$. This establishes identifiability and shows where each baseline fails.

The second experiment is real cross-vendor transfer: train on Siemens only and test on GE/Philips raw data. We will request evaluation on the fastMRI transfer set from its maintainers or obtain GE/Philips raw data from collaborating sites. We will not substitute IXI here: its multi-site data is magnitude-only and cannot test raw operator physics.

The third experiment evaluates hallucination. We insert synthetic lesions and resection-like cavities, which mimic the fastMRI 2020 pseudo-vessel failure, into fully sampled test images before simulating shifted acquisition. We then measure whether our operator-explained versus residual uncertainty map localizes spurious structure, using area under the ROC curve against the known insertion masks. The decisive check has two parts. On real cross-vendor data, operator-only adaptation must recover at least most of AdaDiff's gain over the naive sampler at lower per-scan cost. It must also beat the metadata-harmonization baseline when metadata is withheld. If either fails, the hypothesis that the transfer wall is operator shift is wrong for that vendor pair, which is itself a result worth reporting.