PaperlayerView exact revision on arXiv

Source-linked AI summary

Orientation-Robust Latent Motion Trajectory Learning for Annotation-free Cardiac Phase Detection in Fetal Echocardiography

Yingyu Yang, Qianye Yang, Can Peng, Elena D'Alberti, Olga Patey, Aris T. Papageorghiou, J. Alison Noble

arXiv:2602.06761v2eess.IVcs.CV

TL;DR

Manual ED/ES annotation and variable fetal heart orientation complicate automated cardiac phase detection in fetal 4CV echocardiography. ORBIT uses registration-based self-supervision to learn a latent trajectory of cardiac deformation without annotations, achieving consistent performance across normal and CHD cases and diverse orientations. The study supports annotation-free, orientation-robust phase detection, while subtype-level CHD conclusions remain limited by sample size.

  • Problem

    Existing fetal cardiac phase detection methods commonly require labour-intensive manual ED/ES annotations and face challenges from variable fetal heart orientations.

  • Method

    ORBIT uses registration-based self-supervision and latent trajectory learning to model cardiac deformation and infer ED/ES frames without manual annotations.

  • Results

    ORBIT maintains annotation-free, orientation-robust performance across normal and CHD cases after training exclusively on normal fetal echocardiography videos.

  • Takeaways & Limitations

    ORBIT supports robust cardiac phase detection directly from fetal 4CV echocardiography across diverse fetal orientations.

  • Takeaways & Limitations

    Subtype-level CHD findings remain descriptive because limited and unbalanced sample sizes prevent definitive disease-specific conclusions.

Abstract

from arXiv · show

Fetal echocardiography is essential for detecting congenital heart disease (CHD), facilitating pregnancy management, optimized delivery planning, and timely postnatal interventions. Among standard imaging planes, the four-chamber view (4CV) provides important information for CHD diagnosis, where clinicians carefully inspect the end-diastolic (ED) and end-systolic (ES) phases to evaluate cardiac structure and motion. Automated detection of these cardiac phases is thus a critical component towards fully automated CHD analysis. However, existing approaches typically rely on manual annotation of ED/ES frames, which is labour-intensive and time-consuming. We present ORBIT (Orientation-Robust Beat Inference from Trajectories), a self-supervised framework that identifies cardiac phases without manual annotations under various fetal heart orientation. ORBIT employs registration as self-supervision task and learns a latent motion trajectory of cardiac deformation, whose turning points capture transitions between cardiac relaxation and contraction, enabling accurate and orientation-robust localization of ED and ES frames across diverse fetal positions. Trained exclusively on normal fetal echocardiography videos, ORBIT achieves consistent performance on both normal (mean absolute error 1.9 frames for ED and 1.6 for ES) and CHD cases (mean absolute error 2.4 frames for ED and 2.1 for ES), outperforming existing annotation-free approaches constrained by fixed orientation assumptions. These results highlight the potential of ORBIT to facilitate robust cardiac phase detection directly from 4CV fetal echocardiography.

1. Introduction

Fetal cardiac phase detection is important for automated CHD analysis, but existing methods require costly ED/ES annotations and struggle with variable fetal heart orientations. ORBIT addresses these challenges through annotation-free, orientation-robust self-supervised learning based on latent cardiac motion trajectories.

  • Motivation: Existing fetal ED/ES detection methods rely on manual annotations, while direct unsupervised transfer is challenged by non-standardized imaging windows, fetal heart orientation, and fetal motion.These constraints limit applicability in annotation-scarce settings.
  • Contribution: ORBIT introduces annotation-free, orientation-robust cardiac phase detection for fetal 4CV echocardiography.The framework identifies ED and ES frames without manual labels and is evaluated across fetal orientations, CHD conditions, and preprocessing strategies.
  • Method: ORBIT replaces reconstruction-based self-supervision with registration-based cardiac deformation estimation.Its latent decomposition encodes temporal deformation into a compact trajectory representing cardiac motion rather than image appearance variation.

2. Related work

Cardiac phase detection methods span supervised approaches requiring annotations and unsupervised approaches that exploit cardiac-motion periodicity or low-dimensional representations. In fetal echocardiography, annotation scarcity and orientation variability motivate self-supervised motion modelling that can generalize beyond fixed views.

  • Supervised phase detection uses direct ED/ES regression or indirect cardiac-cycle inference, but depends on substantial annotated data.Reported architectures include CNN-RNNs, Vision Transformers, and fetal echocardiography frameworks combining plane recognition, cycle localization, and abnormal-structure detection.
  • Large unlabelled fetal echocardiography datasets remain difficult to exploit because expert annotations are costly and time-consuming.
  • Unsupervised methods infer ED/ES timing from intrinsic cardiac-motion periodicity without explicit labels, using low-dimensional embeddings or spatio-temporal decompositions.LLE-based embeddings can infer phase transitions through density analysis, while SVD-based approaches model motion from spatio-temporal matrices.
  • Existing unsupervised approaches face consistency, interpretability, and fetal-view limitations, including restriction to apical 4CV and reliance on manual orientation.The earlier LMP framework achieved promising accuracy but was restricted to apical 4CV views and required manual orientation information.
  • The proposed direction learns a latent subspace from temporal deformation fields across complete videos to characterize cardiac motion efficiently and generally.This design is motivated by low-dimensional approaches for representing high-dimensional, nonlinear cardiac motion, including barycentric and polyaffine motion models.

3. Methodology

ORBIT learns cardiac deformation through registration against an abstract reference frame, decomposes it into static and low-dimensional motion components, and analyzes the resulting latent trajectory for annotation-free phase detection.

  • 3.1. Reference-Free Cardiac Deformation Estimation: ORBIT estimates cardiac deformation without a manually specified ED reference by relating each frame to an abstract reference frame.An autoencoder predicts deformation fields that warp observed frames to this intermediate coordinate system, enabling deformation estimation between arbitrary frame pairs.
  • 3.1. Reference-Free Cardiac Deformation Estimation: Stationary velocity fields provide smooth, regular diffeomorphic transformations and efficient bidirectional deformation computation between arbitrary frames.The transformation is parameterized as Φ_t←R = exp(V_t←R), with inter-frame deformation approximated using a first-order Baker–Campbell–Hausdorff formulation.
  • 3.2. Latent Trajectory Learning for Cardiac: The encoder–decoder represents each frame with a latent vector and reconstructs its frame-wise velocity field, separating static structure from motion variation.The static component is shared across the video, while the motion component captures frame-specific deviations around that latent anchor.
  • 3.2. Latent Trajectory Learning for Cardiac: ORBIT constrains motion components to an M-dimensional orthogonal latent subspace and uses learned basis vectors and MLPs to obtain low-dimensional motion coordinates.The latent decomposition is optimized jointly with the network parameters, with M investigated as 1 or 2.
  • 3.2. Latent Trajectory Learning for Cardiac: Unlike LMP’s appearance reconstruction, ORBIT predicts deformation fields through self-supervised registration, optimizing latent coordinates to capture cardiac deformation dynamics.Registration uses normalized cross-correlation between fixed frames and warped moving frames, with each frame registered to subsequent frames in both temporal directions.
  • 3.3. Annotation-Free Cardiac Phase Detection: After training, encoder-derived latent coordinates form a compact trajectory that characterizes periodic cardiac motion and can be analyzed for phase transitions without manual annotations.The trajectory is studied in very low-dimensional motion spaces because cardiac motion is low-dimensional and cyclic.

4. Experiments and settings

The study evaluates ORBIT on normal and CHD fetal echocardiography videos using annotation-free training, orientation analysis, and comparisons with reconstruction-based baselines. Experiments assess detection accuracy, orientation robustness, preprocessing effects, and runtime.

  • 4.1. Datasets: 422 healthy-participant videos form the training and validation data, while testing uses 88 normal videos and 156 CHD videos spanning 13 conditions.Training and validation contain no ED or ES annotations; expert annotations are available for both test sets.
  • 4.2. Implementation: Runtime is measured with all video frames processed jointly in one forward pass, excluding data loading and preprocessing.On normal test videos, total processing takes 18.23 ± 5.74 ms per video on the reported hardware.
  • 4.6. Experiments: Evaluation measures ED/ES mean absolute error against expert annotations while accounting for frame rate, heart rate, and participant-level dependence.Participant-level aggregation is used for nonparametric comparisons because cycles and videos within participants are not independent.
  • 4.5. Baseline methods: ORBIT is compared with LMP and augmentation-matched LMP+rot-aug to isolate registration-based self-supervision from rotation augmentation.LMP requires manual orientation correction, whereas LMP+rot-aug uses the same rotations as ORBIT without correction.
  • 4.6. Experiments: The experiments test normal-data performance, orientation robustness, distributional shift to CHD, and the effect of automatic heart cropping.The model is trained exclusively on normal fetal echocardiography videos, making CHD evaluation an out-of-distribution test.

5.1. In-distribution results

On normal test videos, ORBIT improves cardiac phase detection over LMP and LMP+rot-aug while remaining robust to fetal heart orientation. Registration-based learning produces consistent latent trajectories whose turning points retain a fixed ED/ES interpretation.

  • 5.1.1. Performance on Normal Test Data: ORBIT significantly outperforms LMP and LMP+rot-aug for ED and ES detection, reducing cycle-percentage error by 5% and 7% versus LMP, respectively.Its accuracy approaches inter-observer variability, and LMP+rot-aug has three times ORBIT’s MAE.
  • 5.1.2. Orientation Robustness: Orientation robustness means that the same trajectory turning-point type corresponds to ED or ES regardless of input orientation.This enables phase inference using a fixed peak/valley interpretation without orientation-dependent reassignment.
  • 5.1.2. Orientation Robustness: ORBIT maintains nearly flat ED and ES error across heart orientations, whereas LMP+rot-aug errors increase markedly between 90° and 225°.Figure 5 reports orientation-group MAE in milliseconds, with quantitative values summarized in Table 11.
  • 5.1.2. Orientation Robustness: Across four orientations of one video, ORBIT preserves consistent latent trajectories while LMP+rot-aug inverts one trajectory and swaps ED and ES meanings.The comparison uses M=1 for direct visualization and interpretable trajectory analysis.
  • 5.1.2. Orientation Robustness: Isomap visualization shows orientation-mixed static embeddings for LMP+rot-aug but orientation-grouped static manifolds for ORBIT.The reported disentanglement allows the latent motion component to represent cardiac dynamics more consistently.

5.2. Out-of-distribution results

ORBIT maintains lower cardiac-phase detection errors than LMP and LMP+rot-aug on CHD data, while overall performance remains stable across normal and abnormal cohorts. Subtype findings are exploratory because several diagnostic groups are small, and preprocessing or frame-rate differences affect error interpretation.

  • Performance on Abnormal Test Data: ORBIT achieves lower MAE than LMP and LMP+rot-aug on 156 CHD videos for both frames and cardiac-cycle percentage.Table 4 summarizes the abnormal-test comparison, with all models using M=2.
  • Performance on Abnormal Test Data: Normal-to-abnormal differences are not statistically significant for LMP or ORBIT, indicating comparable accuracy on abnormal fetal hearts.The comparison used participant-level median and mean errors with Mann–Whitney U tests.
  • CHD Subtypes: ORBIT remains stable in VSD but shows larger errors in CAVSD/B and some complex conditions such as HLHS.CAVSD/B ED error reaches 16.9% of the cardiac cycle versus 9.0% in the normal cohort; subtype-level findings are descriptive.
  • CHD Subtypes: Subtype analysis is underpowered, so larger and more balanced CHD cohorts are needed to assess disease-specific performance reliably.The paper cautions that limited participant counts prevent strong statistical conclusions across individual morphologies.
  • Frame Rate: Lower frame rates in CHD data enlarge time-based errors and constrain frame-level inter-observer disagreement, so frame comparisons require caution.Many abnormal cases were recorded at 30 fps, and lower frame rates can reduce apparent frame disagreement without implying easier annotation.
  • Cropping Variations: Automatic cropping increases mean error by approximately one frame on abnormal data across LMP, LMP+rot-aug, and ORBIT.The degradation is attributed to imperfect alignment between automatically extracted and manually defined cardiac regions.

5.3. Influence of Latent Dimension

Latent dimensionality affects the three methods differently. ORBIT is comparatively insensitive to choosing one or two latent dimensions, whereas the baselines show method-specific preferences.

  • LMP: For LMP, M=2 substantially outperforms M=1 on normal and abnormal test sets.The paper attributes this to LMP’s appearance-based learning and sensitivity to nuisance variations such as fetal or probe motion.
  • LMP+rot-aug: For LMP+rot-aug, M=1 consistently outperforms M=2.The constrained one-dimensional space may align cardiac motion across orientations more consistently than a two-dimensional space.
  • ORBIT: ORBIT is less sensitive to latent dimensionality, with M=1 slightly better on normal data and M=2 significantly better for both phases on abnormal data.The larger abnormal-data capacity may represent more complex or atypical cardiac motion.

6. Discussion and Conclusion

ORBIT replaces reconstruction-based self-supervision with inter-frame registration to learn a low-dimensional deformation trajectory for annotation-free ED/ES detection. Its trajectory-based representation remains stable across fetal orientations and performs consistently on normal and CHD data.

  • Discussion and Conclusion: ORBIT infers ED and ES frames without manual annotations by decomposing frame-wise cardiac deformation into a low-dimensional latent motion component.The trajectory captures cyclical contraction–relaxation dynamics from inter-frame registration.
  • Discussion and Conclusion: ORBIT’s deformation-based trajectory is more stable across fetal orientations than reconstruction-based latent-trajectory methods requiring explicit alignment or orientation-dependent interpretation.The model is trained solely on normal fetal echocardiography videos yet performs on both normal and abnormal CHD data.
  • Future Work: Future work will investigate whether ORBIT can capture disease-specific motion signatures for prenatal CHD characterization and detection.This direction is presented as a future possibility rather than a demonstrated result.

A. Congenital heart diseases included in our study

The study provides a complete list of included congenital heart disease conditions in Table 10.

  • Congenital heart diseases included in our study: Table 10 lists the congenital heart disease conditions included in the study.

B. Quantitative results of LMP+rot-aug

Table 11 reports quantitative MAE across heart-orientation groups, comparing MLP+rot-aug with ORBIT (M=1) and complementing Figure 5.

  • Table 11 reports quantitative MAE for each heart-orientation group.
  • The comparison includes MLP+rot-aug and ORBIT with M=1.
  • These results complement the results shown in Figure 5.
Loading 2602.06761v2…