MRI reconstruction networks are trained on data from one scanner and then break on another. Hospitals also cannot freely retrain them. The 2020 fastMRI challenge showed that the best reconstruction networks, mostly "unrolled" designs that alternate a physics step with a learned denoising step, lose substantial quality when moved from Siemens to GE or Philips scanners (Results of the 2020 fastMRI Challenge). Separately, an audit of FDA-cleared devices found that only 1.4% were ever retrained on new data. In a multi-site study, fine-tuning a chest X-ray model on a new hospital recovered up to 0.23 AUC locally but cost the original site an average of 0.176 AUC (Regulating AI Adaptation). A new site therefore needs a fix that works from a handful of its own scans, needs no fully sampled ground-truth images (these are the costly "annotation" of reconstruction), and comes with a checkable statement about how much behavior on the original site has changed.
We propose to keep the trained network exactly as validated and add a small linear correction in its tangent space. The tangent space is the first-order approximation of how the network's output changes when its weights move slightly. Because the corrected model is linear in the correction, three things follow. It can be fitted in closed form from a few target-site scans. It can be supervised only by the scanner's own undersampled measurements. Its change on any stored source-site scan is computable exactly before deployment. Neural tangent kernel (NTK) analysis is the lens: it treats a network near a fixed set of weights as kernel regression on its gradient features. Deep unfolding supplies the physics that makes measurement-only supervision possible. Vendor shift and expert-data scarcity define the regime.
Core Claim
The claim is that, for vendor shift in accelerated MRI, a locked unrolled network plus a closed-form tangent-space correction can do three things. It can recover most of the gap that ordinary fine-tuning recovers. It can do so with fewer target scans and lower variance across random splits. It also gives an exact, pre-deployment bound on source-site output drift, which fine-tuning cannot provide.
Let $f(y;\theta_0)$ be the trained unrolled network mapping undersampled k-space $y$ to an image. Let $J(y)$ be its Jacobian with respect to a chosen subset of $p$ parameters, such as per-stage step sizes, normalization scales, and the last convolution of each denoiser. The deployed model is $f(y;\theta_0)+J(y)\,\delta$, where $\delta$ is a fixed vector the site computes once. At the target site we split each scan's acquired k-space lines into an input set and a held-out set. We fit $\delta$ so that the corrected image, pushed through the known forward operator $A$ (coil sensitivities, Fourier transform, mask), predicts the held-out lines:
$$\delta^\star=\arg\min_\delta \sum_{i\in T}\big\|A_i^{h}\big(f(y_i;\theta_0)+J(y_i)\delta\big)-y_i^{h}\big\|^2+\lambda\sum_{j\in S}\|J(y_j)\delta\|^2+\gamma\|\delta\|^2 .$$
Here $T$ is the set of target scans, $A_i^{h}$ and $y_i^{h}$ are the forward operator and measurements restricted to held-out lines, and $S$ is a stored source calibration set. The middle term is the exact squared change of the deployed model on source scans. The solution is a single $p\times p$ linear solve.
The term $\sum_j\|J(y_j)\delta^\star\|^2$ is an exact drift certificate on the calibration set, and held-out source scans test how well it generalizes. Sweeping $\lambda$ traces an explicit trade-off between target gain and source retention.
Relation to Prior Work
Three NTK papers supply the method but not the setting. NTK-FL builds a global kernel from client Jacobians and solves model updates in closed form across non-IID clients. It is limited to two-layer MLPs on MNIST-scale data. We reuse its closed-form step and its random-projection compression, but we work in parameter space on a deep trained network and add an explicit source-retention term. NTK-MMD uses the tangent kernel as a linear-memory two-sample test. We reuse it only as a diagnostic: distance between source and target tangent features should predict both the transfer drop and the gain from correction. DeepONet-NTK monitors kernel spectra to reweight losses for operator networks on CIFAR-scale images. It neither unrolls a physics solver nor addresses shift, and it uses the kernel to steer training rather than as the deployed object.
One NTK limitation constrains the design. Kernels at random initialization do not learn features, and they drift under long training. We linearize around trained weights, so the features are already learned and only a small correction is asked of the linear model. The key empirical risk is that the vendor gap is too nonlinear for this, and one experiment below tests exactly that.
Unrolled reconstruction papers supply the base networks and the physics. Deep ADMM-Net unrolls ADMM with learned transforms. HUMUS-Net embeds a multi-scale Swin denoiser and reports in-distribution results only. Neither studies vendor transfer. The low-label segmentation methods DyCON and PCL motivate the regime: few labeled volumes, and high variance at two to five subjects. Both operate on reconstructed images. Our supervision instead comes from the measurements themselves, which is possible only because the unrolled network exposes the forward operator. INPUTADAPTER shares our "keep the predictor frozen" stance but changes inputs rather than adding an auditable output correction.
Research Program
The work has three parts.
Building the correction for real unrolled networks. We start with an end-to-end variational network trained on fastMRI multi-coil knee data (Siemens), then repeat with HUMUS-Net. Jacobian-vector and vector-Jacobian products give $J^\top J$ restricted to $p$ parameters, with $p$ between $10^3$ and $10^5$. Random projections, in the style of NTK-FL, compress larger subsets. Inference costs one extra forward-mode pass. We will study which parameter subsets carry the vendor gap: data-consistency step sizes, denoiser normalization, or late convolutions. If the subset is per-stage physics parameters, that would be scientifically informative, because it would localize the shift in the unrolled structure rather than in the learned prior.
Making the certificate meaningful. The calibration bound is exact. To show it transfers, we compare it with measured drift on held-out source scans, in both pixel error and SSIM.
Testing whether tangent-feature distance predicts transfer. We compute NTK-MMD between source and target tangent features and test whether it predicts the size of each shift's drop and the size of the recovered gain. If it does, a site could decide whether correction is warranted before collecting much data.
Baselines
The naive reference point is the locked source network applied zero-shot to the target vendor. This is the transfer-wall condition of the fastMRI challenge. No numeric zero-shot cross-vendor SSIM was found on Lacuna for that track; the challenge reports only that transfer performance was significantly lower than the 4X track's roughly 0.964 SSIM. We will measure the zero-shot result ourselves.
The upper reference is a network trained in-domain on the target vendor with full ground truth. For Stanford 2D knees, HUMUS-Net reports SSIM 0.8954 and E2E-VarNet 0.8928 (HUMUS-Net). These are ceilings, not competitors.
The strongest competitor is ordinary nonlinear fine-tuning on the same few target scans with the same held-out-line self-supervised loss, regularized toward the source weights by an L2 penalty and early stopping, and optionally using a replay buffer of the same source calibration set. It is the strongest because it uses identical information and has more flexibility. It is also the method whose forgetting Regulating AI Adaptation documents. Fairness requires several conditions. It updates the same parameter subset and a full-parameter variant. It sees the same scans and the same source calibration set as replay. It gets an equal hyperparameter search budget (learning rate, steps, penalty weight) selected on a held-out target scan. Each method gets the same wall-clock compute. A low-rank adapter variant is included under the same budget.
Beating it means one of two outcomes at $M\in\{1,2,5,10\}$ target scans with ten random splits. Either equal target SSIM at equal or lower source drift, or a Pareto-dominant target-gain versus source-drift curve, plus lower split-to-split variance. If fine-tuning matches the curve, our remaining contribution is the certificate alone, and we will report it that way.
Two ablations isolate the key ingredients. The first applies the identical objective, optimized nonlinearly through the network, so that only the linearization differs. The second sets $\lambda=0$, so that the source term's role is isolated. A third comparison uses supervised correction with fully sampled target images, which measures what measurement-only supervision costs.
Why not just…?
Why not just fine-tune on the target scans with a penalty toward the source weights? This is the real threat. A weight-space penalty does not control output change on source scans, and the 0.176-AUC forgetting in Regulating AI Adaptation shows how loosely the two are coupled. At two to five subjects, PCL shows how unstable iterative training is: its CHD Dice at $M=2$ is only 0.356. The matched-budget comparison above settles whether accuracy or stability actually differs. Even if accuracy ties, fine-tuning yields no pre-deployment drift number.
Why not just harmonize the acquisition, for example by matching readout oversampling and intensity normalization? The challenge traced part of the transfer wall to such vendor defaults, so this may close much of the gap cheaply. We treat harmonization as a mandatory preprocessing step applied to every method, including zero-shot. We then report only the residual gap. If harmonization closes everything, the result is negative for our method and still worth stating.
Why not just run per-scan test-time adaptation using data consistency? Per-scan optimization changes the deployed weights on every patient. That defeats the locked-model requirement and provides no source certificate. It also costs many backward passes per scan. We include it as an accuracy reference on target scans to see how much the single fixed correction leaves on the table.
Proposed Experiments
The main shift is fastMRI multi-coil knee (Siemens) to the Stanford 2D and 3D FSE knee sets (GE, per their source), at 4X and 8X acceleration. Ground truth is used only for evaluation. A secondary shift is cross-contrast within fastMRI brain. A third is cross-anatomy brain-to-chest compressive-sensing MRI on the dataset of Deep ADMM-Net. On that dataset the original model reached 37.68 dB zero-shot, which gives a published starting point. If access to the challenge's hidden GE and Philips transfer scans can be arranged through its organizers, that is the ideal test. Metrics are SSIM, PSNR, and NMSE on target; drift on held-out source scans; variance across splits; and, for a subset, blinded radiologist hallucination ratings, since the challenge documented hallucinated structures.
Three checks are decisive.
Does linearization hold? If the nonlinear version of our objective clearly beats the linear one on target SSIM at every $M$, the tangent approximation is insufficient for vendor shift. The method then reduces to its certificate.
Does the calibration certificate predict held-out source drift? We require rank correlation and the bound to be within a small factor of measured drift across $\lambda$. Otherwise, the regulatory value is void.
Does tangent-feature NTK-MMD order the three shifts by their zero-shot drop and by their recovered gain?
Failure on the first check while the second holds would still leave an adapter whose effect on validated cases can be audited. Success on all three would give sites a closed-form adapter fitted from a few scans with an auditable bound on source drift.
References
- Results of the 2020 fastMRI Challenge for Machine Learning MR Image Reconstruction
- Regulating AI Adaptation: An Analysis of AI Medical Device Updates
- Neural Tangent Kernel Empowered Federated Learning
- Neural Tangent Kernel Maximum Mean Discrepancy
- A DeepONet joint Neural Tangent Kernel Hybrid Framework for Physics-Informed Inverse Source Problems and Robust Image Reconstruction
- Deep ADMM-Net for Compressive Sensing MRI
- HUMUS-Net: Hybrid Unrolled Multi-scale Network Architecture for Accelerated MRI Reconstruction
- DyCON: Dynamic Uncertainty-aware Consistency and Contrastive Learning for Semi-supervised Medical Image Segmentation
- Positional Contrastive Learning for Volumetric Medical Image Segmentation
- Adapt the Data, Not the Model: Input-Space Adaptation for Frozen Time-Series Predictors