The First Benchmark Dataset For Fetal Radial Echocardiography With Full Temporal Coverage
FeEcho4D is the first large-scale benchmark dataset for fetal radial echocardiography with full 4D(3D+t) temporal coverage, designed to support prenatal cardiac functional assessment and clinical translation. It contains over 1,845 expert-annotated 3D radial volumes from 52 subjects, spanning diverse gestational ages and fetal poses to capture the complexity of real-world imaging. By employing a radial slicing strategy that preserves clinical acquisition geometry and anatomical symmetry, FeEcho4D enables more reliable segmentation, motion estimation, and 3D reconstruction for AI-driven fetal cardiac analysis.
Accurate prenatal cardiac assessment is essential for understanding fetal heart development and improving clinical decision-making, but achieving automated analysis remains challenging due to the scarcity of dynamic annotated data, variations in fetal pose, and the inherent complexity of ultrasound acquisition. FeEcho4D addresses this gap by offering a standardized, clinically aligned dataset with full 4D (3D+t) temporal coverage, enabling research and development of more robust, clinically relevant approaches for fetal cardiac studies.
The FeEcho4D dataset comprises 1,845 3D radial ultrasound volumes collected from 52 unique subjects across various gestational ages. Each volume represents a full cardiac cycle with high temporal resolution, including diverse fetal positions and orientations to reflect real-world clinical variability.
All volumes are radially resampled, aligning the center to the 2D image origin. This creates uniformly sampled S radial slices (H × W) at fixed angular intervals, forming structured 3D volumes for consistent downstream analysis.
Using a custom annotation tool, experts mark points along the left ventricular (LV) myocardium and cavity. Given ultrasound noise, users adjust both spatially and temporally to ensure accuracy, refining adjacent slices for smooth contours. All annotations undergo multi-round review and cardiologist validation.
FeEcho4D generates dense cardiac masks by propagating sparse end-diastolic and end-systolic annotations using B-spline Fourier (BSF) motion model, which blends local and global deformations and regularizes trajectories to capture periodic motion. This ensures temporally coherent, anatomically consistent masks that are manually reviewed for accuracy.
The full 3D+t segmentations are reconstructed into smooth, temporally consistent left ventricular (LV) meshes using Graph Harmonic Deformation (GHD). The meshes are further processed with Differentiable Voxelization & Slicing (DVS), enabling differentiable conversion to voxel grids and slices for precise dynamic metrics and seamless integration with learning-based models.
Each dataset frame includes expert-validated masks (LV, MYO) and metadata (gestational age, cardiac phase indices, acquisition parameters, fetal biometry), enabling research in segmentation, tracking, motion analysis, reconstruction, and clinical functional assessment.
The End-Diastolic Volume (EDV) and End-Systolic Volume (ESV) are derived directly from the endocardial mesh volumes \(V_{\mathrm{ENDO,ED}}\) and \(V_{\mathrm{ENDO,ES}}\). The Stroke Volume (SV) and Ejection Fraction (EF) are computed as:
$$SV = EDV - ESV, \qquad EF = \frac{EDV - ESV}{EDV} \times 100\%.$$
For Global Longitudinal Strain (GLS), the LV epicardial mesh is sliced into \(N\) radial planes aligned with the long axis. Each plane is indexed by \(\theta\), representing its angular position around the long axis. On each plane, the apex-to-base lengths at end-diastole (ED) and end-systole (ES) are denoted as \(L_\theta^{\mathrm{ED}}\) and \(L_\theta^{\mathrm{ES}}\). To minimize the influence of outliers, a subset of the longest and shortest planes is excluded, and the remaining \(N'\) representative planes are used for averaging:
$$L_{\mathrm{ED}} = \frac{1}{N'}\sum_{\theta \notin OL} L_\theta^{\mathrm{ED}}, \qquad L_{\mathrm{ES}} = \frac{1}{N'}\sum_{\theta \notin OL} L_\theta^{\mathrm{ES}}.$$
$$GLS = \frac{L_{\mathrm{ES}} - L_{\mathrm{ED}}}{L_{\mathrm{ED}}} \times 100\%.$$
(In this study, we initially slice \(N=30\) planes indexed by \(\theta\), and retain \(N'=20\) planes after excluding the 5 longest and 5 shortest.)
For Global Circumferential Strain (GCS), the LV epicardial mesh is sampled with \(N\) short-axis cross-sections within the mid-ventricular region (central 10% of the LV long-axis extent). Each cross-section is indexed by \(i\) along the long axis. On each cross-section, circumferences at ED and ES, \(C_i^{\mathrm{ED}}\) and \(C_i^{\mathrm{ES}}\), are computed via ellipse fitting. After excluding several of the longest and shortest circumferences, \(N'\) representative cross-sections remain for averaging:
$$C_{\mathrm{ED}} = \frac{1}{N'}\sum_{i \notin OL} C_i^{\mathrm{ED}}, \qquad C_{\mathrm{ES}} = \frac{1}{N'}\sum_{i \notin OL} C_i^{\mathrm{ES}}.$$
$$GCS = \frac{C_{\mathrm{ES}} - C_{\mathrm{ED}}}{C_{\mathrm{ED}}} \times 100\%.$$
(In this study, we sample \(N=6\) cross-sections indexed by \(i\), and retain \(N'=4\) after excluding the 1 largest and 1 smallest circumferences.)
This process ensures robust strain estimation by averaging over representative planes and cross-sections, while reducing sensitivity to noise and irregular contours.
Coming soon. Code and pre-trained models will be released on our GitHub.
Data access is available upon request for non-commercial research use. Please contact the dataset team or fill in the application form (coming soon).
The FeEcho4D dataset and framework are described in our paper:
"Symmetry-Aware Prompt-Guided Segmentation and Graph-Based Reconstruction for Fetal Cardiac Motion Analysis."
Preprint available soon.
Ph.D.
Department of Bioengineering
Imperial College London
DUT-RUISE
Dalian University of Technology
