Slice Thickness as a Hidden Aging Confound in Perivascular Space Imaging

Enlarged perivascular spaces are fluid-filled channels around small brain vessels. On MRI they appear as thin tubes a few millimeters across, and their number is increasingly read as an imaging sign of the chronic, low-grade inflammation that builds up with age ("inflammaging") and of failing waste clearance in the brain. Counting them automatically is a 3D problem. Each channel is a tube that runs through many slices. Many clinical scans, however, are acquired with thick slices, and the best current segmenters fall back to processing each 2D slice on its own when slices are thick. This discards the depth information that tells a tube apart from a dot or from noise. We propose that this "volumetric context loss" does more than lower accuracy. It produces counting errors that are structured by age, because the geometry of perivascular spaces and of the surrounding tissue changes as the brain atrophies. Population studies that relate perivascular burden to age, blood pressure, or inflammatory genetics may therefore be measuring slice geometry as well as biology.

The proposal has two parts. First, a controlled audit. We simulate thick-slice acquisitions from high-resolution scans where the true 3D channels are known, then measure whether counting error from standard segmenters rises or falls with subject age within a fixed acquisition protocol. Second, a fix. We train a 3D segmenter that takes a thick-slice scan as input and is supervised to reproduce the thin-slice, full-3D channel map. A topology-based counting term keeps each tube as one connected object. The central test is whether the fix removes the age-dependent error, not only whether it raises the Dice overlap score.

Core Claim

When perivascular spaces are segmented slice by slice from anisotropic scans, the resulting lesion counts carry error that correlates with age even after adjusting for scanner and protocol. A model trained to recover out-of-plane continuity from simulated thick-slice inputs removes most of that correlation. The claim is falsifiable in both directions. If the audit finds no age-structured error, the downstream literature is safer than feared and the method contribution shrinks to an accuracy gain. If the fix does not reduce the error, then depth recovery is not the mechanism.

Why This Intersection, and What Already Exists

On the inflammaging side, the file supplies the scientific stakes rather than a method. The VCID workshop proceedings by Lennon et al. (VCID biomarkers workshop; The future of biomarkers for vascular contributions to cognitive impairment and dementia (VCID): proceedings of the 2025 annual workshop of the Albert research institute for white matter and cognition) synthesize evidence that senescent microglia drive neurovascular inflammation that shows up as white matter hyperintensities and enlarged perivascular spaces. They also note that current automated tools suffer volumetric errors. This motivates treating perivascular counts as a readout worth measuring correctly. Shah et al. (cardiovascular ageing; Environmental and genetic predictors of human cardiovascular ageing) built a cardiovascular biological-age gap from UK Biobank cardiac MRI and ECG. They found it is accelerated by hypertension (+1.58 years) and linked to TREM2, a microglial inflammation gene. This gives us a systemic aging variable against which brain perivascular burden can be related. Neither paper touches how the brain measurement itself is made.

The closest precedent for our concern is Schilling et al. (White matter geometry confounds DTI-ALPS). They showed that a popular MRI index of perivascular clearance is driven by crossing fibers, axonal undulation, and vessel deviations of up to 30 degrees from assumed axes, rather than by fluid flow. Their analysis spans young and aging HCP cohorts. That paper constrains our design: it shows that 3D geometry can masquerade as a physiological signal. We carry the same suspicion over to segmentation-based counting, which they did not study.

On the volumetric side, Roth et al. (pancreas segmentation; Deep convolutional networks for pancreas segmentation in CT imaging) documented the basic failure: independent 2D predictions drop structures between adjacent slices. The MICCAI 2024 perivascular space challenge (Wu et al.) is the direct benchmark. The winning team used MedNeXt, a ConvNeXt-style segmenter, with "dynamic dimensionality": 3D convolutions on thin-slice scans and 2D convolutions on thick ones. It reached validation Dice of 0.678 in the basal ganglia and 0.545 in the centrum semiovale. These fell to 0.441 and 0.326 on the held-out Shanghai cohort. The challenge reports overlap only. It does not ask whether errors depend on subject age, and its thick-slice fallback reintroduces exactly the context loss we study.

The WMH challenge (Kuijf et al.) similarly found failures on small lesions across scanners. We borrow two method pieces. P-Count (persistence-based counting) counts lesions from a probability map using persistent homology, which tracks which bright regions stay separate as a threshold is swept. It cut count error on ISBI2015 from 26.86 to 19.14 (supervised). Boundary loss (Kervadec et al.) raised WMH Dice from 0.727 to 0.748. We use P-Count's counting idea inside training rather than only after it.

3D TractFormer (Gao et al.) and Swin-UMamba (Liu et al.) show that efficient volumetric backbones can keep 3D context at full-brain scale. Both were tested only on isotropic or macroscopic targets, and TractFormer only on healthy young adults. We treat them as interchangeable backbones, not as our contribution.

To our knowledge, no work in the file measures whether segmentation error for micro-scale aging lesions is correlated with age. None trains a segmenter to recover thin-slice 3D topology from thick-slice input for this purpose.

Research Program

The audit starts from the challenge scans acquired at thin, near-isotropic resolution, with their expert annotations. We simulate thick-slice acquisitions at 3, 4, and 5 mm, in axial and coronal directions, using a slice-profile blur followed by subsampling. The original annotation stays as ground truth. Because the true 3D channels are known, every error can be attributed to thickness.

For each segmenter and thickness we compute per-subject count error $e_i = \hat N_i - N_i$. Here $\hat N_i$ is the predicted number of perivascular spaces for subject $i$ and $N_i$ is the annotated number. We then fit

$$e_i = \beta_0 + \beta_{\text{age}}\, a_i + \gamma^\top c_i + \varepsilon_i,$$

where $a_i$ is age, $c_i$ collects site, simulated thickness, region, and total brain volume, and $\varepsilon_i$ is residual noise. The quantity of interest is the age-bias slope $\beta_{\text{age}}$. A nonzero slope means the measurement tool itself contributes an age trend. We also estimate the orientation of each annotated channel from its principal axis. This tests the proposed mechanism directly: oblique channels crossing few thick slices should be the ones that fragment or vanish, and orientation should shift with atrophy.

The fix is a 3D segmenter trained on pairs of simulated thick-slice input and thin-slice target. It uses a MedNeXt backbone, so the architecture matches the challenge winner, and it outputs a map on the thin-slice grid. Training uses three terms: the usual Dice and cross-entropy terms, boundary loss, and a differentiable persistence penalty. The persistence penalty compares the number of long-lived connected components in the prediction to the target count, discouraging one tube from being split into many dots. At test time the model runs on real thick-slice scans with no simulation.

Baselines

The naive reference point is the published challenge approach run as-is: MedNeXt with dynamic dimensionality, 2D on thick slices and 3D otherwise, as described by Wu et al.. Its reported numbers are 0.678 and 0.545 validation Dice and 0.441 and 0.326 out-of-distribution Dice (basal ganglia and centrum semiovale). It is the strongest existing method on this exact benchmark. We retrain it from the public recipe on the same 100 training scans and the same simulated augmentations, with the same GPU budget and hyperparameter search. We confirm we roughly reproduce the reported validation Dice before comparing.

A second baseline resamples every thick scan to isotropic spacing by interpolation and applies a fully 3D MedNeXt. A third baseline swaps in a Swin-UMamba backbone (Liu et al.) to check that gains are not backbone-specific. No reported result was found on Lacuna for Swin-UMamba on perivascular spaces.

For counting, all baselines are post-processed both by thresholding and by P-Count. On its own dataset, P-Count's supervised count error was 19.14 versus 26.86 for thresholding; it has no reported result on perivascular spaces.

Ablations isolate the key ingredient. We compare thickness augmentation with native-grid targets, thin-grid targets without the persistence term, and the full model. Beating the baseline means two things. The first is a statistically significant reduction in the magnitude of $\beta_{\text{age}}$ at matched thickness. The second is a lower count error on the held-out Shanghai cohort. Dice gains alone do not count as success.

Why not just…?

Why not just resample to isotropic voxels and run a 3D network? Interpolation creates voxels but not information. A tube that passes between two 5 mm slices leaves no trace to interpolate. This is included as the second baseline. If it matches the full model on $\beta_{\text{age}}$, the thin-grid supervision is unnecessary and we will report that.

Why not just augment training with random slice thicknesses? This is the real threat, because it is cheap and standard. It is exactly our first ablation. The distinguishing hypothesis is that augmentation teaches robustness at the native grid, whereas thin-grid targets force the model to infer out-of-plane extent. If augmentation alone closes the age-bias gap, the method contribution collapses to the audit, and we will say so.

Why not just adjust for scanner and protocol in the downstream regression? Covariates remove average offsets per site. They cannot remove error that varies with age within a protocol, which is what $\beta_{\text{age}}$ measures after conditioning on $c_i$. The audit settles this directly. If $\beta_{\text{age}}$ is zero once protocol is adjusted, covariate adjustment suffices.

Why not just count with P-Count after any segmenter? Post-hoc topology can merge fragments but cannot recover channels the segmenter never detected. Applying P-Count to every baseline separates detection loss from fragmentation.

Proposed Experiments

The first decisive check is the audit on simulated thickness within the challenge cohorts (Edinburgh, SG70 and MACC, and held-out Shanghai), all elderly or at-risk populations. It reports $\beta_{\text{age}}$ per method, thickness, and region, with bootstrap confidence intervals, alongside the orientation analysis of which channels are lost.

The second check is real out-of-distribution transfer on Shanghai, reporting Dice, count error, and $\beta_{\text{age}}$.

The third check is a downstream sensitivity test. The HCP aging subjects used by Schilling et al. span ages 35 to 90, and we would use their thin-slice structural scans if available at sufficient resolution. On these subjects we estimate the slope of perivascular burden against age three ways: from thin-slice reference counts, from baseline counts on simulated thick slices, and from our model on the same thick slices. The fraction of the reference slope recovered is the practical measure of harm and repair.

If UK Biobank brain imaging can be linked to the cardiovascular age gap of Shah et al., we will repeat the comparison for the association between perivascular burden and that age gap. This availability must be confirmed, as the file documents only the cardiac data. The test asks whether a claimed heart–brain inflammaging link changes size depending on how perivascular spaces are counted.

The project succeeds if baselines show a reliable nonzero age-bias slope and the full model shrinks it beyond the augmentation-only ablation. It fails informatively if either condition does not hold.