Training on the Source Hospital's Own Thin Slices to Generalize Across Slice Thickness

When a three-dimensional CT or MRI segmentation model is trained at one hospital and fails at another, the usual explanation is "style": different scanners give different contrast, noise and brightness. This proposal tests a second, largely unmeasured explanation. Clinical volumes are sharp within each slice but coarse between slices, and how coarse depends on the hospital's protocol: slice thickness, slice spacing, and how each slice blurs tissue along the body axis. A model trained on one center's fixed slice profile may learn that profile as a shortcut. When it then sees thicker or differently blurred slices, the small structures that span only a few voxels along the body axis are the first to break. The idea is simple. Train only on one center. From its thinnest acquisitions, synthesize many plausible "other hospitals" by changing the in-plane style and the through-plane slice profile independently, following how scanners actually form images. Always supervise the model with labels on the original thin grid, whatever thickness the input has.

Two sentences for a colleague would be these. Existing single-source domain generalization perturbs 2D image frequencies as if every direction were the same, but in anisotropic 3D scans the axis running across slices carries acquisition signature rather than anatomy. We randomize that axis separately with a physical slice model and decouple the output grid from the input grid, then measure how much of real cross-scanner failure this recovers, especially for small organs.

Core Claim

The claim has two parts, and the second part could be false. The first part is diagnostic. Through-plane acquisition shift explains a substantial and measurable fraction of the cross-scanner performance drop in volumetric abdominal segmentation, and that fraction is concentrated in small, thin structures such as the adrenal glands, pancreas, esophagus and gallbladder. The second part is methodological. Treating the two directions differently gives better zero-shot transfer to unseen scanners than three things: isotropic 3D frequency perturbation, resampling to a common spacing, and two-stage "super-resolve then segment" pipelines. Treating them differently means randomizing in-plane low-frequency style the way 2D methods do, randomizing the through-plane slice profile with a physical blur-and-subsample model, and predicting on a fixed thin output grid. The method gets these gains without touching any target-site data.

Formally, from a source thin-slice volume $x$ with thin-grid label $y$, we generate a training input

$$\tilde{x} = D_{\Delta}\big(h_{\sigma} *_z \, T_{\phi}(x)\big),$$

and train the network to predict $y$ on the original thin grid from $\tilde{x}$. Here $T_{\phi}$ is a random in-plane spectral style transform with parameters $\phi$, applied slice by slice to low in-plane frequencies. $h_{\sigma}$ is a one-dimensional slice-profile kernel of random width and shape, convolved only along the body axis $z$. $D_{\Delta}$ subsamples along $z$ to a random spacing $\Delta$. The form follows the standard picture of thick-slice formation: each reconstructed slice is a weighted average of tissue over its thickness, sampled at the slice spacing.

Where This Sits in Prior Work

The motivation comes from three findings. AMOS built a 600-scan abdominal benchmark from two centers and eight scanner models. It reports a drop of more than 20% when models are evaluated on unseen MRI scanners, but it does not say which part of the acquisition difference causes the drop. That attribution is the gap we target. Pellegrini et al. review neuroimaging studies and show that single-center cohorts systematically inflate accuracy. Slice thickness reduction for lung nodule detection shows that through-plane resolution matters for small targets: converting 5mm to 1mm CT raised scan-level AUC from 0.867 to 0.902. That study is single-center and was never tested across institutions, which is exactly the question we ask.

The method draws on two lines of work. The first is single-source frequency augmentation. FreeSDG and RaffeSDG treat low frequencies as institutional style and high frequencies as anatomy, and they perturb or filter 2D images accordingly. We keep their in-plane logic. The difference is the body axis: there, high frequencies are mostly missing or aliased because of the slice profile, so "high frequency equals anatomy" does not hold. Pseudo multi-source DG splits synthetic styles into pseudo-domains. We borrow this by treating sampled slice profiles as pseudo-domains in one ablation. The second line is the 2.5D design of Wang et al., which uses 2D convolutions early and 3D convolutions once features become near-isotropic. This is our backbone, because it keeps in-plane detail at full resolution. Their study was single-institution, and they explicitly ask whether the split survives scanners with different aspect ratios.

Three papers constrain the design. Frequency-domain adversarial training (VAFT) shows that 3D frequency perturbations hurt far more than 2D ones because they break continuity between slices. On Synapse, UNETR's Dice falls from 74.3% to 32.2% under its attack. This warns that careless 3D perturbation can destroy the anatomy we want to keep. It is also why our through-plane perturbation is limited to physically plausible blur and subsampling, not arbitrary spectral changes. DA-VSR uses the sharp in-plane slices as self-supervision to adapt a super-resolution model at test time. We use the same asymmetry, "in-plane is sharp, so it can teach the body axis," but at training time and for segmentation, with no per-volume optimization at deployment. Roth et al. show that slice-wise 2D predictions fragment along the body axis, with pancreas Dice of 56% before and 68% after 3D smoothing. This is why the model must be natively volumetric rather than a slice-wise method like LTUDA, which its authors leave in 2D.

Research Program

The work has three parts.

The first part is a shift-attribution study. We take source-center volumes, hold out a test split, and apply only the through-plane degradation, at the slice thicknesses and spacings found at the other AMOS scanners. Separately, we apply only in-plane intensity and style changes. We compare each of these drops, organ by organ, with the real drop on true external scans. If through-plane degradation alone reproduces a large share of the real small-organ drop, the diagnostic claim holds. If it does not, the methodological part is less likely to matter, and we would report that.

The second part is the training method above. The backbone is the 2.5D U-Net, with in-plane layers at native resolution. The slice-profile kernels are Gaussian and boxcar families with widths sampled to cover clinically common thicknesses. The style transform randomizes in-plane low-frequency amplitude in the spirit of RaffeSDG's Butterworth filtering. Output is always on the source thin grid. A small consistency loss ties predictions from two different synthetic acquisitions of the same volume together.

The third part is an analysis of where it helps. We stratify results by organ volume and by the gap between target and source slice thickness. We expect the generalization curve, Dice as a function of thickness mismatch, to flatten.

Baselines

The naive reference is a standard volumetric pipeline trained on the source center with routine intensity and geometric augmentation, resampling every input to the source spacing before inference. For orientation, AMOS reports a 3D-UNet in-distribution CT mean Dice of 89.04% and the >20% unseen-scanner MRI drop. We will rerun this ourselves under the single-source split, because no reported result was found on Lacuna for single-source cross-scanner training on AMOS. It runs on the same backbone, epochs and hardware as our method.

The strongest existing single-source method is RaffeSDG. Its best reported number is 0.775 Dice on STARE when trained only on DRIVE, against 0.600 for a plain U-Net and 0.677 for FreeSDG. That is a 2D fundus result, so it is not comparable in absolute terms. We will extend it to 3D in two ways: slice-wise in-plane filtering, and a fully isotropic 3D Butterworth filter. Its filter parameters will be tuned on the same source-only validation split with the same trial budget as ours. The adversarial competitor is VAFT, which reports 75.66% clean Dice and 66.33% under attack for UNETR on Synapse. We run it on our backbone and data.

The two-stage alternative is slice-thickness reduction followed by segmentation, following the STR and DA-VSR designs. The super-resolver is trained on the same source thin slices, and in its DA-VSR variant it is given test-time adaptation steps. We report its extra inference cost.

The ablation that isolates our key ingredient is plain slice-thickness augmentation: the same $h_\sigma$ and $D_\Delta$ degradation, with labels downsampled to the thick input grid and the in-plane style transform turned off. Against it we compare axis-separable perturbation with thin-grid supervision.

Beating a baseline means a statistically significant mean Dice gain on unseen-scanner targets, using a patient-level bootstrap. It also means a gain of at least several Dice points on the small-organ subgroup, without loss on large organs or on the in-distribution test.

Why not just…?

Why not just resample everything to a common spacing, as standard pipelines already do? Resampling makes array shapes match but cannot restore detail that thick slices averaged away. Interpolating a 5mm scan to 1mm gives smooth slices that the source-trained model never saw. The naive baseline measures exactly this, and the attribution study tells us whether the residual gap sits in small organs, as we predict.

Why not just super-resolve the test scan and then segment it? This is a credible competitor, since STR showed real gains for small nodules. But it was single-center, and its authors note they had no true thin-slice ground truth. A super-resolver trained at one center is itself subject to the same center bias, and DA-VSR addresses this only with per-volume test-time optimization. We run both, at matched training data. If the two-stage pipeline matches our single model, the remaining advantage is only inference cost, and we will say so.

Why not just extend RaffeSDG or VAFT to 3D isotropically? VAFT's own results show that isotropic 3D spectral perturbation easily destroys continuity between slices. Isotropic filtering also treats body-axis frequencies as anatomy, which they largely are not in anisotropic scans. The 3D RaffeSDG and VAFT baselines test this directly.

Why not just add slice-thickness augmentation to an existing pipeline? This is the most serious threat, because it shares our through-plane degradation. It differs in two ways: it supervises on the degraded grid, and it does not separate in-plane style from through-plane profile. If it performs as well as our full method on external scanners, the methodological claim fails. The contribution would then reduce to the attribution study plus a recommendation to use that simple augmentation. That is still useful, and we would report it plainly.

Proposed Experiments

The main benchmark is AMOS. We train on scans from one center and a subset of scanners, chosen to be the thinnest-slice source available, and test zero-shot on the remaining scanners and the other center. Results are reported per organ with Dice and average symmetric surface distance, and are stratified by slice-thickness mismatch, using whatever spacing metadata the scans provide. Secondary external targets for the CT model are BTCV and AbdomenCT-1K, which LTUDA used for generalization. A separate MRI-only source run on AMOS tests whether the body-axis effect holds in soft tissue. Synapse and ACDC give a direct comparison with VAFT's reported numbers.

The decisive checks are three. The first asks whether through-plane degradation of held-out source scans reproduces most of the real small-organ drop. The second asks whether the full method beats plain slice-thickness augmentation on unseen scanners, specifically for adrenal glands, pancreas and esophagus. The third asks whether Dice stops falling as target slice thickness departs from the source. Positive answers to all three would show that single-center bias in volumetric imaging is partly a resolution shortcut along the body axis. They would also show that this shortcut can be removed using nothing but the source hospital's own thin slices.