Undoing Scanner Shift with a Hundred Parameters

A segmentation model trained at one hospital often fails at the next. A common reason is that the new hospital's MRI scanner renders the same anatomy with a different brightness response and different sharpness. This proposal tests a narrow claim. Much of that failure can be reversed, without labels and without touching the model, by a tiny learned correction applied to the image before it enters the frozen network. The correction is a smooth, monotone remapping of intensities followed by a gain on each spatial-frequency band, about 30 to 60 numbers per site. It is fit on a handful of unlabeled scans from the new site, so that the frozen model's internal activations look statistically like they did on the training data.

If this works, it changes what "adapting to a new hospital" means. Today the options are full retraining, an expensive robust-training recipe, or tuning adapters inside a large foundation model with labeled target data. The alternative here is a per-site correction small enough for a radiologist to plot and inspect, and to reset. That matters under lifecycle regulation, which Regulating radiology AI medical devices that evolve in their lifecycle describes as lacking clear rules for how much a deployed model may change. Beyond the method, the project asks a measurable scientific question: how many degrees of freedom does a real cross-scanner shift have? We answer it by tracing recovered accuracy against the number of adapted parameters, from 30 up to the full network.

Core Claim

The claim is that cross-scanner shift in MRI segmentation is mostly low-dimensional and mostly lives in two places: the scanner's intensity response and its frequency response. A correction with on the order of $10^2$ parameters should therefore recover most of the accuracy lost on an unseen scanner. "Most" is defined below against a labeled oracle. Such a correction should beat simple intensity harmonization, and it should match adapter-based tuning that uses thousands to millions of parameters.

The correction is

$$x' = \mathcal{F}^{-1}\big(g_\theta(|k|)\cdot \mathcal{F}(h_\phi(x))\big),$$

where the symbols are as follows. $x$ is the target-site volume. $h_\phi$ is a monotone piecewise-linear intensity map with $K$ knots ($K\approx 16$). $\mathcal{F}$ is the 3D Fourier transform, computed with physical voxel spacing so that anisotropic slices are handled correctly. $|k|$ is radial spatial frequency, and $g_\theta$ is a positive gain shared within each of $B$ radial bands ($B\approx 16$). The gain is real and positive, so it never changes the phase of the spectrum.

Leaving phase alone is a deliberate design choice taken from FedDG. That work exchanges amplitude spectra between hospitals on the premise that scanner "style" lives in amplitude while anatomy lives in phase. We reuse that decomposition, but as a fitted per-site correction rather than as training-time augmentation.

The parameters are fit with a frozen source model $f$. For each normalization layer $l$, we store the per-channel activation mean $\mu_l^{s}$ and standard deviation $\sigma_l^{s}$ computed on source data. These summary statistics are the only thing that leaves the source hospital. We then minimize

$$\mathcal{L}(\phi,\theta)=\sum_l \big\|\mu_l(x')-\mu_l^{s}\big\|^2+\big\|\log\sigma_l(x')-\log\sigma_l^{s}\big\|^2+\beta\,R(\phi,\theta),$$

where $\mu_l(x')$ and $\sigma_l(x')$ are the same statistics computed on corrected target volumes, and $R$ pulls the correction toward the identity map.

Relation to Prior Work

The motivating evidence is that the gap is large and is not closed by standard preprocessing. AMOS is a two-center benchmark spanning eight scanner models. It reports that models lose over 20% mean Dice on MRI from scanners unseen in training, despite standard normalization. Do Vision Foundation Models Enhance Domain Generalization in Medical Image Segmentation? shows a vanilla U-Net collapsing to 1.22 Dice under a spine CT-to-MRI shift.

Existing remedies fall into three groups, and none does what we propose. Training-time robustness methods change how the source model is trained. These include FreeSDG, which mixes frequency bands and adds a second, reconstruction decoder, and Domain Game, which uses two encoders to separate anatomy from scanner noise. They work from a single source site but are heavy, and they cannot repair a model that has already been validated and deployed. Multi-source methods require several labeled hospitals at training time. FedDG shares amplitude banks across hospitals, and DCAC generates convolution weights from a predicted domain code.

The third group adapts parameters inside the model. Cekmeceli et al. tune Rein, LoRA, or Ladder adapters inside frozen 2D foundation models, with labels. PEMMA tunes 8% of a UNETR to add PET, and 3DSAM-adapter tunes 16.96% of SAM for prompted 3D tumors. All of these spend thousands to millions of parameters inside the network and need target labels or prompts.

Our method is closest in spirit to Adapt the Data, Not the Model, which adapts the inputs of frozen time-series predictors. We bring that idea to 3D imaging with a physically motivated, phase-preserving parameterization. We also make the parameter count itself the object of study.

Two cited results constrain the design. FASS shows that high-frequency texture separates low-contrast pancreatic lesions from normal tissue. Similarly, S2M-Net concedes that its fixed low-frequency truncation risks discarding diagnostic detail. Our band gains can boost or damp each frequency band but never truncate it, and the experiments measure small-structure accuracy separately to catch any loss of detail.

Research Program

We first train frozen source models on a single site. Two backbones are used. The first is a standard 3D U-Net in the nnU-Net style, the strong conventional model. The second is HNOSeg-XS, a 3D model built on frequency-domain operators with fewer than 34.7k parameters. HNOSeg-XS lost only 3.5% Dice under resolution change, against 14.4% for nnFormer, but it was never tested across scanners. This pairing lets us ask a secondary question: does a compact frequency-domain model respond to input-spectrum correction differently from a convolutional one?

The central analysis is a parameter-count curve. On each target site we adapt progressively larger subsets: the input correction (about 30 to 60 parameters), the normalization scale and shift parameters (thousands), LoRA-style adapters (around 1%), and finally the full network. Each subset is fit under two regimes. In the unlabeled regime, the subset is fit with the feature-statistics objective. In the oracle regime, it is fit with Dice loss on target labels.

The two regimes separate two questions. The oracle regime asks whether the correction family is expressive enough to undo the shift at all. The unlabeled regime asks whether label-free matching finds that correction. If the oracle-fit input correction recovers most of the labeled full fine-tuning gain, the low-dimensionality claim holds regardless of how well the unlabeled objective performs.

We also inspect the fitted corrections. For each scanner we plot $h_\phi$ and $g_\theta$. We then test whether corrections cluster by vendor and field strength, and whether a correction fit on one scanner transfers to another scanner of the same model. Clustering of this kind would be direct evidence that the method captures hardware response rather than noise.

Baselines

The naive reference point is the frozen source model applied with only its standard per-volume normalization. Every other number is reported as the fraction of the gap closed between this baseline and labeled full fine-tuning on the target site.

The simplest serious competitor is classical intensity harmonization: histogram matching of each target volume to a stored source intensity histogram. This is exactly our method with the band gains fixed to 1 and $h_\phi$ chosen by matching rather than learned. It therefore doubles as the ablation that isolates the frequency term. Beating it means a statistically significant Dice gain, by paired test across target volumes, from adding the band gains.

The strongest label-free competitor at test time is test-time adaptation of the network's normalization parameters. It uses the identical feature-statistics objective, the same unlabeled target volumes, and the same number of optimization steps. It has no citation in the file, but it is the natural objection. Beating it, or matching it with roughly 100 times fewer parameters and no change to the network, counts as success.

A second label-free competitor follows FedDG's mechanism at test time: swap the low-frequency amplitude of each target volume with an average source amplitude, without learning. It is fit-free, so its cost is reported separately.

Among training-time competitors, we retrain the source model with FreeSDG-style frequency-mixed augmentation on the same single source site. This asks whether robust training removes the need for any adaptation. We also include Domain Game, whose reported prostate result is an out-of-distribution drop limited to approximately 4.7% versus 9–15% for prior single-source methods. The file gives relative figures only, so we rerun it under our split.

For context on what multi-source access buys, the reported multi-site prostate figures are FedDG at 87.39% average Dice (FedAvg 85.57%) and DCAC at 90.51% Dice on Domain 2. These use multiple labeled source sites, so they are a ceiling for a single-source method rather than a direct comparison. Matching them would be a strong result.

The labeled competitor is adapter tuning with five labeled target volumes, under the same step budget. We use Rein-LoRA and Ladder as in Cekmeceli et al., applied to the 3D U-Net, and labeled full fine-tuning on the same five volumes.

Why not just…?

Why not just histogram-match target scans to the source? It is the right first thing to try, and it is built in as the ablation. The case against it as a full solution is AMOS's over-20% drop on unseen MRI scanners after standard normalization. Scanners also differ in blur and noise texture, which a pointwise intensity map cannot represent. If the band gains add nothing over matching on prostate and AMOS MRI, we will report that the shift is essentially an intensity-response effect. The low-parameter claim would then survive in an even simpler form.

Why not just adapt the normalization layers at test time? This is the most serious threat, because it uses the same objective and is equally cheap to run. Our bet is that correcting the input once beats rescaling every layer. The input correction has fewer degrees of freedom, so it should overfit less to five unlabeled volumes, and it leaves the validated network bit-identical. The head-to-head comparison at matched data and steps settles this. If normalization adaptation wins clearly, the publishable finding becomes the parameter-count curve, and the input-correction claim is dropped.

Why not just train a robust source model once, with frequency augmentation or disentanglement? FreeSDG and Domain Game report real gains. However, they leave residual drops; Domain Game, the best case, still loses about 4.7%. They also cannot help a model already in service. We test both whether our correction helps on top of a robust source model and whether it closes the gap from a plain one. Complementary gains would be the practical message.

Why not just tune adapters in a foundation model with a few labels? Cekmeceli et al. show this generalizes well in 2D. It needs labels at each site, though, along with large frozen backbones and per-site changes to the network's weights. The oracle-fit input correction will show how much of the labeled adapter gain about 50 parameters can capture. If adapters remain far ahead even under oracle fitting, the shift is not low-dimensional and our core claim is refuted.

Proposed Experiments

The primary benchmark is the six-site T2 prostate MRI collection used by FedDG, DCAC, and Domain Game. We train on each single site and test on the remaining five, using 5, 10, and 20 unlabeled target volumes for fitting.

The second benchmark is AMOS MRI, split by scanner model, with organs grouped by size so that small structures such as the adrenal glands and duodenum are reported separately. The third is BraTS, with BraTS-Africa as a lower-field target, following Domain Game.

Metrics are Dice and average surface distance per structure, the gap-closed fraction defined above, the number of adapted parameters, and adaptation wall-clock time.

Three checks are decisive. First, the oracle-fit input correction should close at least half of the labeled full fine-tuning gap on prostate and AMOS MRI. This is the low-dimensionality test. Second, the unlabeled correction should beat histogram matching and amplitude swap, and should match test-time normalization adaptation within its confidence interval. Third, small-structure and lesion Dice should not fall below the no-adaptation baseline, which guards against the high-frequency hazard raised by FASS.

A failure on the first check falsifies the central claim. A failure on the third check alone would point to a regularizer that keeps the high-frequency band gains near 1.