Latent Cardiac Motion Trajectories

• Latent Cardiac Motion Trajectories are learned, time-indexed paths that compress high-dimensional cardiac motion into structured, clinically relevant representations.
• Techniques such as structure-motion decomposition, Fourier parameterization, and latent ODE models enable precise phase detection and motion reconstruction in imaging modalities.
• These models offer actionable insights by facilitating simulation of heart cycles, enhancing prognostic evaluations, and driving improved clinical decision making.

Searching arXiv for the papers on arXiv to ground the article in current literature. {"query":"Latent cardiac motion trajectories echocardiography latent motion profiling ORBIT latent ODE cardiac MRI Fourier motion modelling", "max_results": 10} {"query":"arXiv (Yang et al., 7 Jul 2025) Latent Motion Profiling for Annotation-free Cardiac Phase Detection in Adult and Fetal Echocardiography Videos", "max_results": 5} {"query":"arXiv (Yang et al., 6 Feb 2026) Orientation-Robust Latent Motion Trajectory Learning for Annotation-free Cardiac Phase Detection in Fetal Echocardiography", "max_results": 5} {"query":"arXiv (Lan et al., 2 Jun 2026) Conditional Latent Diffusion Model with Fourier-based Motion Modelling for Virtual Population Synthesis", "max_results": 5} {"query":"arXiv (Brüggemann et al., 25 Jun 2026) A Latent ODE Approach to Spatiotemporal Modeling of Cine Cardiac MRI", "max_results": 5} Latent cardiac motion trajectories are time-indexed paths in learned representation spaces that encode how cardiac anatomy evolves through a beat. In current cardiac imaging literature, the latent object may be a low-dimensional subspace trajectory extracted from echocardiography video reconstruction, a Fourier-parameterized path in a mesh-VAE latent space, a continuous neural-ODE state over a normalized cine CMR cycle, a probabilistic motion matrix decoded to diffeomorphic deformations, or a graph-structured latent state on anatomical regions (Yang et al., 7 Jul 2025, Lan et al., 2 Jun 2026, Brüggemann et al., 25 Jun 2026, Krebs et al., 2020). Across these formulations, the central goal is consistent: compress spatiotemporal cardiac motion into a representation that preserves cycle structure, supports reconstruction or generation, and exposes clinically useful landmarks such as end-diastole, end-systole, strain trajectories, or risk-associated remodeling patterns.

1. Conceptual scope and representational principles

A recurrent design choice is the separation of motion from non-motion factors. In Latent Motion Profiling (LMP), each frame latent is decomposed as a structure-motion sum, with a static anatomy component $z^s$ and a dynamic motion component constrained to a low-rank orthogonal subspace; in the reported setting, $K=2$ produces interpretable axes aligned with septal and lateral wall motion (Yang et al., 7 Jul 2025). ORBIT adopts an analogous split in fetal echocardiography, but the decoder outputs stationary velocity fields (SVFs) for registration rather than reconstructed frames, so the latent motion trajectory is defined by deformation coordinates rather than appearance variation (Yang et al., 6 Feb 2026). In free-breathing ungated cardiac MRI, a different factorization separates VAE-derived motion signals from a synthetic contrast latent, enabling contrast-only or motion-only image synthesis (Zou et al., 2022). In label-free ultrasound synthesis, MAFE explicitly separates appearance features from motion features, with pseudo re-identification guidance applied to appearance and pseudo optical flow guidance applied to motion (Li et al., 10 Dec 2025).

This common decomposition reflects a broader methodological claim: cardiac motion is treated as a low-dimensional, structured, and often repeatable component embedded within much higher-dimensional image or mesh observations. The literature does not enforce one canonical latent format. Some papers use explicit framewise coordinates $\{a_t\}$ in a small Euclidean subspace, some use global sequence codes, some use coefficient sets that generate the full cycle, and some use latent node states over anatomical graphs. This suggests that “latent cardiac motion trajectory” denotes a family of representations rather than a single architecture or objective.

2. Mathematical parameterizations of latent motion

One widely used formulation is a structure-motion decomposition. LMP defines

$z_t = z^s + z_t^m,\qquad z_t^m = E a_t,$

where $E=[e_1,\dots,e_K]\in\mathbb{R}^{D\times K}$ is an orthogonal motion basis and $a_t\in\mathbb{R}^K$ are time-varying motion coordinates. The latent trajectory is therefore the sequence $\{a_t\}_{t=1}^T$ in a $K$-dimensional motion subspace; with $K=2$, the trajectory forms a smooth back-and-forth loop whose extremities align with end-systole and end-diastole (Yang et al., 7 Jul 2025). ORBIT retains the same high-level factorization but maps the latent sum to an SVF and then to a diffeomorphic deformation, using

$$\hat z_t = z^s + z_t^m,\qquad \Phi_{j\leftarrow i}=\exp(V_{j\leftarrow R}-V_{i\leftarrow R}),$$

so that the trajectory is tied directly to deformation states rather than to image reconstruction (Yang et al., 6 Feb 2026).

A second formulation parameterizes the entire cycle analytically. In 4D F-MeshLDM, the latent trajectory of a subject’s 3D heart anatomy is written as a truncated Fourier series over normalized phase $K=2$0:

$K=2$1

The paper uses $K=2$2, corresponding to $K=2$3 non-DC harmonics and $K=2$4 Fourier terms. Because the Fourier basis is $K=2$5-periodic, $K=2$6, so cycle closure is exact in latent space by construction (Lan et al., 2 Jun 2026).

A third formulation is continuous-time latent dynamics. The cine CMR latent ODE model defines

$K=2$7

with heart-rate-aware phase warping $K=2$8 learned through a positive function $K=2$9. Here the latent state evolves continuously over one normalized cardiac cycle, and the decoder reconstructs biventricular meshes at arbitrary phases (Brüggemann et al., 25 Jun 2026).

A fourth formulation treats the full cycle as a latent matrix. “Learning a Generative Motion Model from Image Sequences based on a Latent Motion Matrix” defines a motion matrix $\{a_t\}$0 whose columns parameterize deformations from the end-diastolic frame to each target frame, while each row is a latent trajectory for one motion mode. The prior is a zero-mean multivariate Gaussian with temporal Gaussian-process structure, and the decoder maps each column to a stationary velocity field and then to a diffeomorphism $\{a_t\}$1 (Krebs et al., 2020).

These parameterizations differ chiefly in where periodicity and smoothness are imposed. In low-rank subspace models, periodicity may emerge from reconstruction or registration pressure. In Fourier models, it is enforced analytically. In neural ODE and GP models, it is mediated by continuous-time dynamics or covariance structure rather than by explicit frame indexing alone.

3. Learning objectives, priors, and self-supervision

The supervisory signal for latent cardiac motion trajectories ranges from pure reconstruction to registration, stochastic dynamics, and biomechanics-derived priors. LMP is trained with a reconstruction loss that combines a Fréchet-mean image term for the static anatomy latent and a framewise reconstruction term for the full latent. The paper explicitly states that no temporal, cycle-consistency, contrastive, or explicit periodicity losses are used; periodic and interpretable motion patterns emerge from the low-rank motion subspace and reconstruction pressures (Yang et al., 7 Jul 2025). ORBIT replaces reconstruction with registration-based self-supervision, summing normalized cross-correlation terms over temporal offsets $\{a_t\}$2 in both directions, while generating diffeomorphic deformations through exponential maps of SVFs (Yang et al., 6 Feb 2026).

Other models strengthen temporal regularity more explicitly. DeepTag uses a bidirectional generative diffeomorphic registration network to infer inter-frame motion and a differentiable composition layer to accumulate those motions into Lagrangian trajectories relative to end-diastole, together with a global sequence-level similarity constraint (Ye et al., 2021). GPTrack introduces a sequential Gaussian Process in latent space and bidirectional recursive aggregation, reporting that the combination promotes temporal consistency and spatial variability of cardiac dynamics (Yang et al., 2024). A different prior source appears in the biomechanics-informed temporal VAE, which is trained not on cine images but on deformation sequences generated by finite element simulations; at test time, motion tracking proceeds by latent space exploration under image similarity and a weak Gaussian prior (Qin et al., 2022).

These differences matter because they determine what the latent trajectory is expected to preserve. Reconstruction-only models emphasize compactness and interpretability. Registration-based models emphasize deformation fidelity. GP and ODE models emphasize temporal continuity. Simulation-trained priors emphasize biomechanical plausibility. The literature therefore treats the latent trajectory both as a compressed representation and as an inductive bias on admissible cardiac motion.

4. Annotation-free phase detection in echocardiography

The most direct clinical use of latent cardiac motion trajectories in echocardiography is annotation-free detection of end-diastole (ED) and end-systole (ES). LMP learns a 2D latent motion trajectory from 4-chamber echocardiography, estimates a principal orientation by RANSAC and PCA, projects the 2D trajectory to a 1D signal, applies Savitzky–Golay smoothing and optional high-pass filtering, and then assigns ES to peaks and ED to valleys (Yang et al., 7 Jul 2025). On the EchoNet-Dynamic test set of 1,276 videos, the unsupervised method reports ED 3.0 (3.3) frames = 58.3 (66.4) ms and ES 2.0 (2.2) frames = 39.8 (43.0) ms; the abstract separately reports ES as 38.8 ms. On fetal echocardiography, transfer learning from the adult model yields match-pair MAE ED 1.46 (1.26) frames = 20.7 (17.9) ms and ES 1.74 (1.48) frames = 25.3 (21.7) ms, with reported detection coverage of 95% ED frames and 96% ES frames (Yang et al., 7 Jul 2025).

ORBIT addresses a more specific limitation of fetal echocardiography: large variability in fetal position and acquisition angle. It uses registration as the self-supervision signal, manual heart-region cropping, temporal downsampling during training, and orientation augmentation by rotating each video by $\{a_t\}$3, $\{a_t\}$4, $\{a_t\}$5, and $\{a_t\}$6 (Yang et al., 6 Feb 2026). Turning points are detected on a 1D latent motion sequence using scipy’s find_peaks, with a fixed validation-derived assignment of peaks versus valleys to ED and ES. Trained exclusively on normal fetal echocardiography videos, ORBIT reports MAE $\{a_t\}$7 frames for ED and $\{a_t\}$8 for ES on normal cases, and MAE $\{a_t\}$9 frames for ED and $z_t = z^s + z_t^m,\qquad z_t^m = E a_t,$0 for ES on CHD cases, outperforming annotation-free approaches constrained by fixed orientation assumptions (Yang et al., 6 Feb 2026).

Taken together, these studies establish two distinct mechanisms for phase discovery from latent trajectories. In LMP, ED and ES emerge as geometric extrema of a reconstructed latent path. In ORBIT, they emerge as turning points of a deformation trajectory learned through registration. A common misconception is that latent phase detection in fetal ultrasound requires a fixed apical four-chamber orientation. The reported ORBIT results do not support that claim, although they also show that severe CHD such as HLHS and CAVSD/B can blur turning points and increase phase error (Yang et al., 6 Feb 2026).

5. MRI motion tracking, strain estimation, and latent trajectory recovery

In cardiac MRI, latent cardiac motion trajectories are often tied to diffeomorphic displacement fields, Lagrangian motion, or deformation-aware latent features rather than to direct phase labels. DeepTag uses a bidirectional generative diffeomorphic registration framework on cardiac tagging MRI, where the latent variable is a stationary velocity field that generates inter-frame motion and a differentiable composition layer accumulates those motions into Lagrangian trajectories relative to the reference frame. On the reported test set, DeepTag achieves RMS $z_t = z^s + z_t^m,\qquad z_t^m = E a_t,$1 mm, with $z_t = z^s + z_t^m,\qquad z_t^m = E a_t,$2 pixels having non-positive Jacobian determinant, while outperforming HARP, OF-TV, VM, and VM-DIF (Ye et al., 2021).

LaMoD brings diffusion modeling into this setting. It extracts deformation-based shape features from cine CMR using the encoder of a pre-trained LDDMM-based registration network, applies latent diffusion under DENSE supervision, and decodes refined latent motion features into transformation fields and strain maps. The paper reports that LaMoD significantly improves the accuracy of motion analysis in standard CMR images and improves myocardial strain analysis, with visual and boxplot comparisons showing reduced displacement and segmental circumferential strain errors relative to StrainNet, UNetR, and 3D TransUNet (Xing et al., 2024).

GPTrack emphasizes temporal and spatial structure jointly. It combines bidirectional recurrence with Gaussian Process latent coding and reports average Dice $z_t = z^s + z_t^m,\qquad z_t^m = E a_t,$3 on CardiacUDA and $z_t = z^s + z_t^m,\qquad z_t^m = E a_t,$4 on ACDC, together with very low percentages of non-positive Jacobians, while remaining computationally efficient (Yang et al., 2024). LAPANet addresses a different part of the pipeline by estimating non-rigid motion directly from accelerated k-space rather than from reconstructed images, accommodating as few as $z_t = z^s + z_t^m,\qquad z_t^m = E a_t,$5 Cartesian lines/frame or $z_t = z^s + z_t^m,\qquad z_t^m = E a_t,$6 radial spokes/frame and achieving temporal sampling below $z_t = z^s + z_t^m,\qquad z_t^m = E a_t,$7 ms (Ghoul et al., 2024). A related free-breathing ungated GRE inversion recovery framework learns cardiac and respiratory motion latents from centered k-space via a VAE and then excites those latents to synthesize contrast-only series for $z_t = z^s + z_t^m,\qquad z_t^m = E a_t,$8 mapping or motion-only series for breath-hold CINE generation (Zou et al., 2022).

These MRI studies show that the latent trajectory is frequently not the final clinical object. Instead, it is an intermediate state from which one derives dense motion fields, Lagrangian paths, strain tensors, or reconstructed cine volumes. This suggests that, in MRI, the importance of latent cardiac motion trajectories lies as much in their ability to regularize physically plausible deformation as in their ability to compress temporal information.

6. Generative modeling, virtual populations, and motion-conditioned synthesis

Generative models use latent cardiac motion trajectories both as internal state representations and as controllable objects for simulation. 4D F-MeshLDM learns a mesh-VAE latent space, fits each subject’s motion with Fourier coefficients, and trains a conditional diffusion prior over those coefficient tokens. Because the motion trajectory is defined by a truncated Fourier series, cycle closure is exact in latent space. On 5,000 UK Biobank subjects, the model reports volume cycle consistency $z_t = z^s + z_t^m,\qquad z_t^m = E a_t,$9, mesh cycle consistency $E=[e_1,\dots,e_K]\in\mathbb{R}^{D\times K}$0, conditional BiV coverage@5 $E=[e_1,\dots,e_K]\in\mathbb{R}^{D\times K}$1, and Seq. RMSE $E=[e_1,\dots,e_K]\in\mathbb{R}^{D\times K}$2 mm (Lan et al., 2 Jun 2026).

RePCM addresses single-frame bi-ventricular motion completion. It first learns vertex-wise motion descriptors and derives a motion-based regional partition, then uses a Region-Specific Injection Module together with a phenotype-adaptive mixture-of-experts prior conditioned on ED shape to synthesize the full cycle from one ED mesh. Across ACDC, M&Ms, and M&Ms-2, it reports LV ASSD $E=[e_1,\dots,e_K]\in\mathbb{R}^{D\times K}$3 and RV ASSD $E=[e_1,\dots,e_K]\in\mathbb{R}^{D\times K}$4, with best results at $E=[e_1,\dots,e_K]\in\mathbb{R}^{D\times K}$5 regions and $E=[e_1,\dots,e_K]\in\mathbb{R}^{D\times K}$6 experts (Yang et al., 20 May 2026). Here the latent motion trajectory is not a framewise coordinate curve but the decoded full-cycle displacement field generated from a static latent code and ED anatomy.

Motion-conditioned video diffusion adopts a different role for the latent trajectory. In MFD-V2V, an LTMA registration network learns latent motion states that decode to stationary velocity fields and displacement fields, and a spatio-temporal motion encoder converts those fields into motion features for conditioning a video diffusion model. The method reports FID/KID/FVD/FID-VID of $E=[e_1,\dots,e_K]\in\mathbb{R}^{D\times K}$7, and downstream LV segmentation Dice improves from approximately $E=[e_1,\dots,e_K]\in\mathbb{R}^{D\times K}$8 on raw DENSE to approximately $E=[e_1,\dots,e_K]\in\mathbb{R}^{D\times K}$9 on synthesized cine (Deb et al., 1 Jul 2025). CardioDiT removes explicit factorization and models the whole 3D+t cine CMR volume jointly in a 4D latent diffusion transformer, reporting FID $a_t\in\mathbb{R}^K$0 on public validation and EF Wasserstein-2 distance $a_t\in\mathbb{R}^K$1, together with improved inter-slice consistency and realistic cardiac function distributions (Seyfarth et al., 26 Mar 2026). In echocardiography, MCDM conditions latent video diffusion on a single global motion vector $a_t\in\mathbb{R}^K$2 extracted from bidirectional ED$a_t\in\mathbb{R}^K$3ES motion features; for 64-frame generation it reports FID $a_t\in\mathbb{R}^K$4 and FVD16 $a_t\in\mathbb{R}^K$5 (Li et al., 10 Dec 2025).

These models differ in whether the latent cardiac motion trajectory is itself generated, decoded from a static code, or used as conditioning for appearance synthesis. The shared pattern is that motion is treated as a structured prior on admissible cardiac videos or meshes rather than as a by-product of image generation.

7. Prognostic significance, graph formulations, and unresolved issues

Latent cardiac motion trajectories have also been used as prognostic phenotypes. The latent ODE cine CMR model studied 72,386 UK Biobank participants without baseline cardiovascular disease and used a covariate-conditioned prior on the end-diastolic latent state together with heart-rate-aware continuous dynamics. In held-out evaluation, adding the latent score to refitted pooled cohort equations improved the stratified C-index from $a_t\in\mathbb{R}^K$6 to $a_t\in\mathbb{R}^K$7, compared with $a_t\in\mathbb{R}^K$8 for seven established cardiac markers (Brüggemann et al., 25 Jun 2026). Earlier, 4Dsurvival learned a Cox-supervised latent code from standardized right-ventricular motion trajectories in 302 patients with pulmonary hypertension and achieved an optimism-corrected Harrell’s C-index of $a_t\in\mathbb{R}^K$9 (95\% CI: $\{a_t\}_{t=1}^T$0–$\{a_t\}_{t=1}^T$1), versus $\{a_t\}_{t=1}^T$2 (95\% CI: $\{a_t\}_{t=1}^T$3–$\{a_t\}_{t=1}^T$4) for a benchmark based on conventional RV volumetrics (Bello et al., 2018).

Another strand of work represents the trajectory on graphs rather than in Euclidean latent vectors. The spatiotemporal graph neural process represents cardiac dynamics as a spatiotemporal multiplex graph, models latent trajectories with a GNN-parameterized neural ODE, and infers uncertainty over latent initial states and control variables from sparse context observations. It reports, on ACDC, thickness MAE $\{a_t\}_{t=1}^T$5 mm and volume MAE $\{a_t\}_{t=1}^T$6 ml for interpolation, up to $\{a_t\}_{t=1}^T$7 accuracy for 5-class classification, and on UK Biobank up to $\{a_t\}_{t=1}^T$8 accuracy for atrial fibrillation detection (Banus et al., 16 Sep 2025). This does not replace earlier latent trajectory formulations; rather, it relocates them onto anatomical relational structure.

Several limitations recur across the literature. Exact periodicity is not universal: 4D F-MeshLDM guarantees cycle closure analytically, whereas the latent ODE model explicitly states that a cycle-consistency penalty did not help and therefore omits it (Lan et al., 2 Jun 2026, Brüggemann et al., 25 Jun 2026). Orientation robustness is not automatic in fetal ultrasound: LMP relies on aligned apical 4CH data, while ORBIT was designed to address apical, transverse, and basal variability through registration supervision and rotation augmentation (Yang et al., 7 Jul 2025, Yang et al., 6 Feb 2026). Generalization to pathology remains incomplete. Reported open problems include arrhythmias, severe motion artifacts, and non-apical views for LMP; severe CHD cases such as HLHS and CAVSD/B for ORBIT; fixed-topology constraints in mesh generative models; and the need for external validation in more representative patient cohorts for prognostic latent ODE models (Yang et al., 7 Jul 2025, Yang et al., 6 Feb 2026, Lan et al., 2 Jun 2026, Brüggemann et al., 25 Jun 2026).

The literature therefore supports a constrained conclusion. Latent cardiac motion trajectories are now a unifying abstraction across echocardiography, cine CMR, tagging MRI, mesh synthesis, and motion-conditioned diffusion. Yet the abstraction remains method-dependent: some trajectories are explicit loops with geometric extrema, some are coefficient sets with exact closure, some are continuous latent states with quasi-periodicity, and some are graph-structured regional dynamics. This suggests that the enduring value of the concept lies less in a single formal definition than in its capacity to organize cardiac motion as a learned, structured, and clinically actionable dynamical representation.