When the Motion Tracker Inherits the Anatomy: Measurement-Induced Pleiotropy in Self-Supervised Cardiac Phenotypes Across Ancestries

Several methods now recover 3D motion from ordinary 2D image sequences without any 3D labels. They start from a template shape, deform it, project the result back into 2D, and penalize mismatch against the apparent motion between frames (optical flow) computed by an off-the-shelf model. These methods are about to become a source of new traits for genetics. A cine-MRI scan, a short movie of the beating heart taken as a stack of 2D slices, can be turned into a dense field of how each piece of muscle stretches and contracts, which is far richer than the single number "ejection fraction." This note argues that the template these pipelines start from is a hidden confounder. Wall thickness and chamber shape differ between people and are partly heritable. If a mismatch between the template and the person's anatomy distorts the recovered motion, then genetic variants that shape anatomy will look as if they shape motion.

Average anatomy also differs across ancestral groups. The same template-driven distortion will therefore be larger in some populations than in others. In a cross-ancestry causal analysis it would look exactly like the "pleiotropic shift" that statistical geneticists already blame on population differences: a genetic variant appearing to act through a second biological pathway in one population but not another. We propose to measure this artifact, to show whether it changes cross-ancestry genetic conclusions, and to remove it with a subject-specific grounding procedure. Correctness is checked through the self-supervision signal itself and through genetic diagnostics, not through 3D labels that do not exist.

Core Claim

Template-based unsupervised motion grounding produces measurement error that depends on anatomy. Because anatomy is heritable and its distribution varies by ancestry, this error creates two problems. It adds spurious genetic correlation between static shape and recovered motion, which we call measurement-induced pleiotropy. It also creates spurious differences in genetic effects across ancestries for the motion traits. Both can be detected without 3D ground truth, and both should shrink when the template is replaced by geometry fitted to each subject's own first frame.

What would be new is the following. Nobody has asked whether failures of unsupervised grounding are structured by population, and nobody has asked whether ancestry-bias diagnostics in genetics can be fooled by an upstream imaging artifact.

Where This Sits

The grounding side supplies the method and also the motivating weakness. Mao et al. supervise 3D occupancy flow by comparing rendered projections to 2D flow and masks. SplatFlow distills motion from the 2D flow model SEA-RAFT into a continuous 4D motion field. RigAnyFace aligns 3D mesh deformations to RAFT displacement fields through differentiable rendering. We reuse this recipe of 2D flow as teacher and differentiable projection as the bridge, here a differentiable slicer for MRI planes.

The weakness comes from CloDS. Its dynamics learning "breaks down if initial geometry is distorted." RigAnyFace adds sensitivity to mesh discretization and to thin shell-like geometry, which is a reasonable description of thin myocardial walls. All of these papers treat initialization error as a robustness issue measured on average. None asks whether it is systematically linked to a property of the subject.

Prior cardiac motion work, Mesh-based 3D Motion Tracking in Cardiac MRI using Deep Learning and Diagnosing Cardiac Deformations using 3d Optical Flow, recovers myocardial deformation for diagnosis. It contains no genetics and no population-stratified evaluation, so it serves as the application domain rather than a competitor. Quality-control work on the UK Biobank cardiac study (Large-scale Quality Control of Cardiac Imaging in Population Studies, Automated Quality Control in Image Segmentation) shows that per-image failure detection is a known concern at biobank scale. That work targets segmentation failures, not anatomy-correlated bias in motion that would otherwise pass QC.

The genetics side supplies the diagnostics and the stakes. GSMR uses its HEIDI-outlier filter, which flags variants whose effects disagree with the others, to remove pleiotropic instruments. Removing 16 APOE-region variants erased a spurious LDL-cholesterol to Alzheimer's link. Our point is that HEIDI-style filters assume pleiotropy is biological, whereas measurement-induced pleiotropy can be shared across many anatomy-associated instruments and so evade outlier tests.

The file lists exactly this blind spot as an open problem: population-specific pleiotropy can survive MR-Egger, HEIDI and MR-PRESSO when instruments share directional confounders. LOG-TRAM estimates local cross-ancestry genetic correlation in 1 Mb windows. It would faithfully report divergence, but it cannot tell biological divergence from divergence caused by measurement, so it is a downstream consumer whose inputs we try to clean. Invariant Ancestry Search concerns causal ancestors in a graph, not human ancestry. Its idea that true causal relations stay invariant across environments motivates our use of ancestry groups as environments for a stability check on the anatomy-to-motion relation.

Research Program

We first build two grounding pipelines on the UK Biobank cardiac cine-MRI cohort. They are identical in architecture and differ only in where geometry comes from. Both use a 2D flow teacher (RAFT or SEA-RAFT, as in the cited work) on each long-axis and short-axis slice sequence. Both deform a 3D left-ventricular mesh, slice it differentiably at the acquired planes, and minimize the gap between projected and observed 2D motion.

The template variant initializes from a single population-mean mesh scaled to the subject. The subject-specific variant fits the mesh to that subject's own end-diastolic frame before any motion is learned, following CloDS's first-frame initialization but with a fitting loss that penalizes wall-thickness mismatch explicitly.

From each pipeline we extract regional peak strain and strain-rate trajectories as the motion phenotypes. We extract wall thickness and chamber volume as the static anatomy phenotypes.

Second, we construct a synthetic testbed where 3D truth exists. We generate left ventricles across a controlled range of wall thicknesses and shapes, assign them known motion fields, and render them into cine-like slice stacks at UK Biobank resolution. This lets us measure exactly how strain error depends on anatomy for each pipeline. The prediction is that the template pipeline shows error that rises with distance from the template, and the subject-specific pipeline shows a much flatter curve.

Third, we develop label-free diagnostics for real data. The first is the per-subject reprojection residual, meaning the unexplained 2D flow after fitting, stratified by anatomy and by genetic ancestry group. The second is the partial correlation between recovered strain and static anatomy after conditioning on clinically established covariates.

Fourth, we run the genetic tests. These are SNP heritability and genetic correlation between anatomy and motion under each pipeline. They also include a bidirectional MR of anatomy on motion with GSMR, and a cross-ancestry comparison of motion-trait effect sizes between the European subset and the non-European UK Biobank imaging participants.

The non-European imaging subsets are small. The genetic tests of contamination are therefore powered mainly in the European subset. The ancestry-dependent claims rest on the synthetic curve combined with measured differences in anatomy distribution, plus effect-size concordance at the strongest loci. We will state this limitation plainly rather than oversell trans-ancestry GWAS on thin samples.

Baselines

The naive reference point uses segmentation-derived clinical indices. Ejection fraction and segmentation-based global strain are computed from the same UK Biobank scans, using the segmentation that the cited quality-control studies audit. This baseline asks whether dense self-supervised motion adds genetic signal at all. We consider it beaten if the subject-specific phenotypes show higher heritability, or loci not tagged by ejection fraction, at matched sample size.

The strongest competitor is the template-based self-supervised pipeline. It represents the published design pattern of CloDS, RigAnyFace and SplatFlow, transplanted to cardiac slices. It gets the same flow teacher, mesh resolution, optimization steps and hyperparameter search budget as our variant; only initialization differs.

No reported result was found on Lacuna for self-supervised cardiac strain accuracy or strain heritability. The nearest reported template-versus-learned contrast is in facial rigging: RigAnyFace's 1.01 mm MAE against NFR's 2.77 mm on facial meshes (RigAnyFace). That shows template choice can move error by a factor of nearly three, but it is not a cardiac number. Beating this competitor means three things, all required. Synthetic strain error must be lower and flatter across anatomy. The anatomy–motion genetic correlation must be smaller, with no loss of motion-trait heritability. Cross-ancestry effect concordance at top loci must be at least as high.

Two ablations isolate the key ingredient. An ancestry-group template variant fits one template per genetic ancestry group. It tests whether coarse group-level correction suffices or whether the subject-level fit matters. A flow-teacher swap exchanges RAFT and SEA-RAFT. It tests whether the artifact comes from the flow teacher rather than the geometry.

For the genetic side, the reference method is GSMR with HEIDI filtering, run identically on phenotypes from both pipelines. LOG-TRAM serves as the cross-ancestry correlation estimator wherever sample size permits. For scale, the file reports LOG-TRAM's 49% average gain in effective sample size and 17–24% PRS $R^2$ gain in WeGene on 29 static traits (LOG-TRAM). These figures are context for what local cross-ancestry modeling can achieve, not a target for motion traits.

Why not just…?

Why not just use ejection fraction and other clinical scalars? These scalars sidestep template bias by discarding regional dynamics, and they are the honest default. However, the file's evidence from other domains shows that coarse summaries miss real motion structure. SplatFlow outperforms 3D-box-based StreetGS in reconstruction fidelity, with 33.64 PSNR against 32.46 for PVG on Waymo (SplatFlow). That is a vision result, not a cardiac one. The naive baseline comparison will show directly whether dense phenotypes add genetic information. If they do not, the contamination finding still matters as a warning to groups adopting these pipelines.

Why not just regress out wall thickness and volume as covariates? This is the serious threat. If template distortion is roughly linear in anatomy, covariate adjustment could remove most of it. It is also cheaper than refitting meshes. But adjusting for a heritable trait that lies on a causal path to cardiac function can introduce collider bias. It can remove real biology along with the artifact, and it fails if the distortion is nonlinear.

We run covariate-adjusted template phenotypes as an explicit arm. The synthetic testbed tells us whether the distortion is linear in anatomy. The genetic arm tells us whether adjustment cuts motion-trait heritability more than subject-specific fitting does. If adjustment matches subject-specific fitting on both, we will report that the cheap fix suffices.

Why not just run LOG-TRAM or MR-PRESSO on the template phenotypes and let them handle cross-ancestry divergence? These tools model genetic architecture, not measurement. LOG-TRAM would record an artifact-driven local divergence as real. Outlier-based pleiotropy filters assume pleiotropic variants are a minority. The file notes that such filters miss pleiotropy shared through directional confounders, and anatomy-driven distortion is exactly a shared directional confounder. The comparison of GSMR/HEIDI outputs between the two pipelines tests this directly.

Why not just fit one template per ancestry group? This requires ancestry labels and treats a continuous anatomical spread as a few categories. It may help on average while leaving the within-group anatomy dependence intact. It is included as an ablation, and if it matches subject-specific fitting we will say so.

Proposed Experiments

The first decisive check is synthetic. Across the generated anatomy range we plot strain error against wall-thickness deviation from the template for each pipeline. The core claim fails if the template pipeline's error is flat in anatomy.

The second check is label-free on real UK Biobank scans. If the reprojection residual and the strain–anatomy partial correlation both differ between template and subject-specific pipelines, and differ across ancestry groups in the direction predicted by their anatomy distributions, the artifact is present in vivo.

The third check is genetic. Under the template pipeline, we expect higher anatomy–motion genetic correlation and a GSMR anatomy-to-strain causal estimate that survives HEIDI filtering. Under subject-specific fitting, we expect both to shrink while motion-trait heritability is preserved. The covariate-adjustment arm is evaluated against the same three outcomes.

The final check is cross-ancestry. For the strongest motion-trait loci, we compare the concordance of effect sizes between European and non-European imaging participants under each pipeline. Higher concordance after subject-specific fitting would show that part of the apparent ancestry-specific pleiotropy was measurement, not biology. Given the small non-European samples, we pre-register this as a directional consistency test rather than a discovery claim.