Doctors can now measure how each region of the heart muscle stretches and squeezes over a heartbeat. This is called myocardial strain, and it is extracted by software that follows tissue across video frames of an MRI or ultrasound scan. It catches damage that the standard "fraction of blood pumped" number misses. A strain value only means something, though, when it is compared with healthy hearts measured the same way, and hospitals do not measure the same way. A scanner that records fewer frames per heartbeat blurs the contraction peak in time. Thick imaging slices blur motion in space. Different tracking software exaggerates wall thickening differently. A strain biomarker found in 100 patients at one hospital therefore tends to fall apart at the next.
We propose to treat these acquisition differences as what they physically are: a small blurring-and-scaling operation applied to motion, not an arbitrary offset added to each number. The plan has three parts. We map every heart's strain over the cycle onto shared anatomical coordinates. We build a normative healthy atlas once, from a large cohort. Then, at each new site, we fit a handful of operator parameters from as few as 5 to 20 local healthy controls and apply that operator to the atlas. Patients are then compared with norms "as this scanner would have measured them." The atlas is moved to the site, rather than the site's data being un-blurred back to the atlas.
Core Claim
Frame rate, slice thickness and tracker choice mostly act on myocardial motion as temporal smoothing with a phase shift, plus direction- and depth-dependent gain. The clearest example of the gain is the radial overestimation by feature tracking, "up to 100%" relative to tagging, in Global and regional left ventricular myocardial deformation measures by magnetic resonance feature tracking in healthy volunteers.
Location-and-scale harmonization of the ComBat family cannot represent a shift in the timing of peak strain or a flattened peak. It also treats spatial locations as exchangeable, so it can break the physical coupling between neighbouring regions. We therefore fit a low-dimensional forward operator from a few local healthy controls, pooling evidence across atlas locations by empirical-Bayes shrinkage. We apply it to a normative atlas and score patients as deviations from the site-matched norm.
We claim this approach should recover disease signal under leave-one-site-out testing better than two alternatives: per-feature ComBat on the same atlas features, and ComBat on clinical 17-segment strain. The advantage should be largest when the new site's control sample is tiny.
Where this sits relative to prior work
Tracking methods (supply the tracking). Three lines of work produce the dense 3D motion we need, and all three were developed and evaluated on the uniform UK Biobank protocol:
- MulViMotion fuses short- and long-axis cine MRI into 3D motion.
- DeepMesh deforms one template mesh per subject, which gives vertex correspondence, but was tested only on healthy UK Biobank subjects. - Biomechanics-informed Neural Networks regularize 2D registration with a simulation-trained prior and transfer to the multi-disease ACDC dataset.
None of them asks how their strain outputs shift between acquisition protocols. We reuse them as fixed trackers.
Anatomical coordinates (supply the coordinates). An Automated Computational Pipeline for Generating Large-Scale Cohorts of Patient-Specific Ventricular Models computes Laplace–Dirichlet ventricular coordinates, meaning depth through the wall, base-to-apex position and angle around the ventricle, on static meshes. It reports that these coordinates fail on malformed hearts and at the outflow tract, which constrains us to exclude the basal outflow region. We attach time-resolved strain to these coordinates, which that pipeline never did.
Harmonization with strict site-held-out testing (supplies the methodology). Using brain structural neuroimaging measures to predict psychosis onset combines ComBat, age and sex modelling and a held-out site, reporting 85% internal versus 73% external accuracy. It harmonizes static brain morphometry, not signals over time.
Evidence that motivates the problem. Batch Effects In Brain Foundation Model Embeddings shows that large pretrained representations keep site artifacts. Deep Learning-Derived Myocardial Strain replaces speckle tracking with contour tracking and still reaches only an inter-vendor ICC of 0.45, with a −2.33% bias against 3D strain. That paper attacks vendor variance upstream in the tracker. We attack it downstream, at the level of the norm.
Physics priors (supply the constraints). Learning-Based Regularization for Cardiac Strain Analysis via Domain Adaptation and Flow Network Tracking impose near-incompressibility and a closed loop over the cycle inside tracking. We use the same constraints as checks that harmonization must not violate.
Research Program
Representation. For each subject, a fixed tracker produces 3D displacement over the cycle. Displacements are resampled to a common normalized cycle of $T$ phases and to atlas nodes in ventricular coordinates. Circumferential, longitudinal and radial strain are then derived from the displacements at each node. The normative atlas is the mean and covariance of this field over UK Biobank subjects, with age and sex modelled additively as in ComBat-GAM.
Site operator. For site $s$, observed strain in direction $d$ at wall depth $\rho$ is modelled as
$$\tilde\varepsilon_{s,d}(x,t) = g_{s,d}(\rho)\,\big(k_{s} * \varepsilon_{d}(x,\cdot)\big)(t-\tau_s) + \eta,$$
where each term is defined as follows:
- $\varepsilon_d(x,t)$ is reference strain at atlas node $x$ and cycle phase $t$.
- $k_s$ is a circular low-pass kernel. It is circular because the cycle is periodic, and one bandwidth parameter captures frame-rate and temporal-filtering effects.
- $\tau_s$ is a phase shift.
- $g_{s,d}(\rho)$ is a smooth gain that varies across wall depth. It is expressed in two to three basis functions per strain direction, and it captures radial exaggeration and through-plane blurring.
- $\eta$ is residual noise.
That is roughly a dozen parameters per site. They are estimated from local healthy controls by maximizing likelihood under the atlas covariance, with a hierarchical prior shrinking site parameters toward those of previously seen sites. The shrinkage is what should make 5 to 20 controls sufficient. A patient is scored by the whitened deviation of their field from the operator-transformed atlas.
Why apply the operator forward. Applying the operator to the atlas avoids deconvolving noisy patient data. Blurring a periodic field stays periodic. As a safeguard, we also check that the implied displacement fields keep myocardial volume change within the range trackers report; MulViMotion reports 8.62%.
Confounding safeguard. Site parameters are fit only on controls, never on patients. This addresses the known failure mode in which a site that scans mostly patients has its disease signal removed as "batch."
Baselines
Pooled, unharmonized features. This is the do-nothing reference. The same atlas deviation scores are computed against the raw UK Biobank atlas, and classifiers use identical features and tuning. Beating it is necessary but not impressive.
ComBat on the 17 clinical segments. ComBat, the empirical-Bayes location/scale correction that borrows strength across features, is applied to peak strain in the standard 17-segment parcellation. It is fit on the same local controls, and the classifier budget matches. It is the strongest simple competitor because it is what a clinical study would actually do. There is no reported result on Lacuna for ComBat applied to cardiac strain. The nearest reference numbers are the neuroimaging 85% to 73% internal-to-external drop in the ENIGMA psychosis study and the 0.866 AUC for circumferential strain in single-center STEMI/NSTEMI discrimination in Quantitative evaluation of acute myocardial infarction by feature-tracking CMR.
Per-node ComBat on the full atlas field. This is the obvious combination of the input directions and the closest competitor to our method. Location and scale are fit independently at every node and phase, using the same controls. It isolates whether the operator structure matters beyond having the atlas.
Ablations of our own method. We run three:
- the operator without the temporal kernel and phase shift (gain only);
- the operator without shrinkage across sites;
- the operator fit on 17 segments instead of atlas nodes.
What counts as winning. Our method wins if, at 10 local controls, it beats per-node ComBat on held-out-site AUC by a margin whose bootstrap interval excludes zero. It must also reduce error in recovering peak-strain timing and radial peak on the controlled benchmark below. If it matches ComBat on 17 segments only once controls reach 20 or more, we will report that as the regime boundary.
Why not just…?
Why not just ComBat the 17 segment values? This is the real threat. Segment averaging already smooths out much spatial noise, and ComBat is well understood. Its weakness is structural: it adjusts means and variances of peak values and cannot model a site that flattens and delays the strain curve. Any biomarker that uses timing or curve shape, such as post-systolic shortening or diastolic recovery rate, is invisible to it. The controlled benchmark measures this directly by comparing timing-error recovery. If segment ComBat matches us on peak-only biomarkers, we will say so and restrict our claim to shape and timing features.
Why not just train the tracker with acquisition augmentation, such as random frame dropping and slice thickening? This attacks the problem upstream, as Deep Learning-Derived Myocardial Strain did. That paper's residual inter-vendor ICC of 0.45 suggests upstream fixes leave a lot behind. Augmentation also cannot recover temporal information that a low-frame-rate scan never captured, so information is lost before tracking. We will train an augmented version of the tracker as an extra baseline. If it removes the gap, the operator is unnecessary. Our method also stacks on top of it, so the fair test is whether the combination still helps.
Why not just use global longitudinal strain or ejection fraction? Scalars are the most robust to site effects, but they discard exactly the regional and directional information that separates infarct severity. In the AMI feature-tracking study, circumferential strain reached AUC 0.866 against 0.788 for longitudinal strain. We include a global-longitudinal-strain classifier with its own harmonization as a floor. If regional atlas features do not beat it out-of-site, the atlas adds nothing clinically.
Proposed Experiments
Controlled acquisition shift with known ground truth. From UK Biobank cine MRI, where the 50-phase cycle is the reference, we synthesize "sites" by re-acquiring the same subjects' images:
- temporal subsampling to 25 and 20 phases;
- slice thickening;
- added noise;
- swapping the tracker (MulViMotion versus a 2D registration tracker).
The true reference strain of each subject is known, so we can measure how accurately each method maps healthy controls from a new site back onto norms. The metrics are root-mean-square error of the whole strain curve, peak-timing error, radial peak error and loop-closure error. We report them as a function of the number of local controls: 5, 10, 20 and 50.
Real site transfer under small samples. Norms come from UK Biobank. The test site is ACDC, which spans several scanner field strengths and five diagnostic groups. The operator is fit only on ACDC's healthy group, holding out controls by cross-fitting, and patients are classified from deviation scores. We compare MI versus normal, DCM versus normal and HCM versus normal against the baselines above. To bring this closer to the small-discovery setting, classifier features are selected on at most 100 subjects from one acquisition condition and tested on another. The reference point for tracker quality on this data is the 0.783 registration Dice on ACDC reported by Biomechanics-informed Neural Networks. We report no strain-classification comparison number from prior work because none appears on Lacuna.
Decisive check on confounding. On ACDC we deliberately fit the site operator on a mix that includes patients. If our method is behaving correctly, disease AUC should drop under that contamination and recover when the operator is fit on controls only. This shows that the method removes acquisition effects and not pathology.
Secondary physics check. Harmonized norms must keep implied volume change and loop closure within tracker-reported ranges. Per-node ComBat is expected to violate these; if it does not, the physics argument is weaker than we think, and we will say so.
Scope limit. Paired multi-vendor echo data, such as the 43-patient GE/Philips repeat scans, would be the natural ultrasound extension. The file gives no public access route to it, so it is out of scope for this proposal.
References
- Global and regional left ventricular myocardial deformation measures by magnetic resonance feature tracking in healthy volunteers: comparison with tagging and relevance of gender
- MulViMotion: Shape-Aware 3D Myocardial Motion Tracking From Multi-View Cardiac MRI
- DeepMesh: Mesh-Based Cardiac Motion Tracking Using Deep Learning
- Biomechanics-Informed Neural Networks for Myocardial Motion Tracking in MRI
- An automated computational pipeline for generating large-scale cohorts of patient-specific ventricular models in electromechanical in silico trials
- Using brain structural neuroimaging measures to predict psychosis onset for individuals at clinical high-risk
- Batch Effects In Brain Foundation Model Embeddings
- Deep Learning-Derived Myocardial Strain
- Learning-Based Regularization for Cardiac Strain Analysis via Domain Adaptation
- Flow network tracking for spatiotemporal and periodic point matching: Applied to cardiac motion analysis
- Quantitative evaluation of acute myocardial infarction by feature-tracking cardiac magnetic resonance imaging