🚧 Under Development ⚠

🫀Official Repository for

4D Reconstruction of Fetal Left Ventricle from Echocardiography via 2.5D Radial Segmentation and Graph-Fourier Reconstruction

FeEcho4D: The First Benchmark 4D Fetal Echocardiography Dataset with 4D Left Ventricular Meshes

Authors:
Md. Kamrul Hasan, Qifeng Wang, Haziq Shahard, Lucas Iijima, Nida Ruseckaite, Yihao Luo, Iris Scharnreitner, Andreas Tulzer, Bin Liu, Guang Yang, Choon Hwai Yap

---

Overview

This repository provides a complete pipeline for 4D fetal cardiac reconstruction from echocardiography, using a three-stage framework for left ventricular (LV) analysis:

• FeEcho4D – Radial data preparation: extract 2D slices by rotating planes around the LV center (Block A)
• SCOPE-Net – Geometry-aware segmentation: prompt-guided, symmetry-consistent segmentation on radial slices (Block B)
• Graph-Fourier Mesh Reconstruction – High-fidelity 3D reconstruction: generate temporally consistent LV meshes from sparse radial slices using GHD + DVS, enabling clinical metric estimation (Block C)

---

📦(A) FeEcho4D Dataset Preparation: Radial Slicing to 4D Mesh Reconstruction

A detailed description of dataset preparation is available on our dataset website 👉 FeEcho4D. This section documents the code used for data preparation. Implementation details and visualizations (overlay images & GIFs) are provided in the folder: (1)Volume-to-Slice

Radial Slicing

Uniform Mask Processing

This step refines raw binary masks into anatomically consistent shapes:

1. Largest component extraction
2. Skeletonization + tail extrapolation
3. Dilation by estimated half-thickness
4. Optional smoothing via binary opening

---

🌀 (B) SCOPE-Net: Symmetry-Consistent Prompt-Enhanced Network

SCOPE-Net is a geometry-adaptive network that explicitly encodes angular symmetry as an inductive prior, unlike fully data-driven deep networks, and integrates spatial prompts through gated modulation of encoder features.

🧭 SCOPE-Net Features

• Flip-Consistent Radial Attention (FCRA) – models angular symmetry in radial views
• Geometry-Aware Self-Supervision – enforces representation-level consistency via an Inter-Slice Augmentation Invariance (ISAI) objective; enables label-free self-distillation and improves feature robustness under strong augmentations
• Prompt Conditioning – supports bounding box or scribble inputs
• Efficient 2.5D Training – 56G FLOPs per frame (vs. 79G for 3D networks)

Implementation – see (2)DL_Segmentation folder for full code details

---

🕸 (C) GHD-based 3D Mesh Reconstruction

Given a sequence of 3D segmentation volumes, the pipeline reconstructs a continuous left-ventricle (LV) mesh by Graph Harmonic Deformation (GHD):

1. Initialize a canonical template mesh M_0 (e.g., sphere or averaged LV shape)
2. Embed vertices {v_i} into a graph structure with Laplacian basis functions
3. Load voxel-wise segmentation masks (binary myocardium/ventricle) and anisotropic voxel spacing
4. Voxelize & Sample – extract point clouds from the mask boundary at each time t
5. Fit – deform M_0 to match sampled boundary points using a GHD energy function:
   • Data term – enforces point-to-surface alignment
   • Regularizers – maintain mesh smoothness and shape consistency
6. Optimize deformation coefficients in harmonic space (low-dimensional basis) for efficiency
7. Iterate over all time frames to produce a smooth temporal mesh sequence {M_t}
8. Export reconstructed meshes as .obj files under each case directory

Implementation – see (3)Slice-to-Mesh folder for full code details

Step 1: Quickstart via Jupyter Notebook

# Step into the (3)Slice-to-Mesh folder
cd /path/to/(3)Slice-to-Mesh

# Launch the notebook for interactive fitting
jupyter notebook ghd_fit_quickstart.ipynb

Step 2: Advanced Execution via Python Script

# Explore ghd_fit.py for full parameter control and customization
python ghd_fit.py \
    --data_root data_example \
    --cases FeEcho4D_017 \
    --times time001-010 \
    --device cuda:0 \
    --mesh_out meshes_out \
    --myo_idx 2

More details can be found in the GHDHeart project.

---

📊 Results: Utility of 3D Reconstruction via Radial Slicing Vs. Standards 2D Vs. Volumetric segmentation

Comparison of SAX, LAX, and Radial slicing (followed by GHD fit) vs. direct 3D volumetric segmentation. Metrics are averaged across LV endocardium (ENDO) and epicardium (EPI). Radial view segmentation consistently outperforms alternatives (p < 0.05 for most cases).

Method	View	HD95 ENDO ↓	HD95 EPI ↓	DSC ↑	MASD ↓	Avg. 2D DSC ↑
UNet	3D Volume	6.99 ± 7.41	5.87 ± 4.65	0.875 ± 0.047	2.30 ± 0.82	0.756 ± 0.286
	SAX	8.95 ± 6.03	5.56 ± 3.11	0.837 ± 0.062	2.15 ± 0.79	0.738 ± 0.267
	LAX	12.68 ± 10.39	6.51 ± 4.65	0.848 ± 0.057	2.72 ± 1.18	0.686 ± 0.287
	Radial	5.57 ± 6.17	4.20 ± 4.58	0.908 ± 0.034	2.08 ± 0.99	0.908 ± 0.041
SAM	3D Volume	6.07 ± 3.05	4.93 ± 2.35	0.874 ± 0.046	2.24 ± 0.73	0.759 ± 0.284
	SAX	5.98 ± 4.07	2.73 ± 1.05	0.874 ± 0.043	1.54 ± 0.49	0.791 ± 0.230
	LAX	7.93 ± 4.44	3.42 ± 1.78	0.876 ± 0.039	2.02 ± 0.62	0.710 ± 0.270
	Radial	4.63 ± 3.44	3.00 ± 2.47	0.917 ± 0.033	1.82 ± 0.67	0.917 ± 0.038

✅ FLOPs/Frame: 79G for 3D, 51G for 2D.

• Improved 2D segmentation accuracy – Radial slicing significantly increases the mean DSC across all image slices compared to SAX and LAX, due to the standardized U-shaped myocardial appearance in all slices.
• Superior 3D reconstruction quality – Radial views outperform SAX, LAX, and volumetric segmentation for both networks, with statistically significant differences (p < 0.05, t-test) in most comparisons.
• Anatomical fidelity & robustness – Higher reconstruction accuracy is attributed to better 2D segmentations and the shape constraints inherent in radial slicing.
• Efficiency gains – 3D reconstruction from sparse 2D radial slices requires fewer FLOPs per frame than direct 3D volumetric segmentation, improving computational efficiency.

Comparison of SAM, UNet, and SCOPE-Net on our FeEcho4D dataset and the public MITEA dataset. SCOPE-Net consistently outperforms baselines in both fetal and adult datasets (p < 0.05 for most metrics).

Dataset	Method	HD95 MYO ↓	HD95 ENDO ↓	HD95 EPI ↓	DSC ↑	MASD ↓	PIA (%) ↓
FeEcho4D	SAM	8.08 ± 3.06	6.06 ± 2.97	5.24 ± 2.97	0.881 ± 0.040	4.71 ± 1.92	0.0258 ± 0.2171
	UNet	13.09 ± 10.26	10.08 ± 9.33	9.31 ± 9.36	0.838 ± 0.140	6.01 ± 3.20	1.39 ± 4.61
	SCOPE-Net	7.46 ± 3.31	5.40 ± 3.13	4.83 ± 3.14	0.893 ± 0.044	4.29 ± 1.71	0.0036 ± 0.0917
MITEA	SAM	6.49 ± 2.20	4.63 ± 3.44	3.00 ± 2.47	0.917 ± 0.033	1.82 ± 0.67	0.0499 ± 0.0602
	UNet	7.30 ± 3.55	5.57 ± 6.17	4.20 ± 4.58	0.908 ± 0.034	2.08 ± 0.99	0.7486 ± 2.35
	SCOPE-Net	5.92 ± 2.37	3.96 ± 2.48	2.48 ± 1.95	0.921 ± 0.033	1.54 ± 0.37	4.5e-4 ± 1.6e-2

✅ FLOPs/Frame: 63 G for SAM, 51 G for 2D UNet, 56 G for SCOPE-Net.

• Consistent superiority – SCOPE-Net significantly outperforms UNet and SAM on FeEcho4D and MITEA across all subregions and metrics (HD95, MASD, DSC, PIA), with p < 0.05 in most cases.
• Boundary & overlap gains – Produces sharper, anatomically faithful segmentations, particularly in challenging regions (LV apex, lateral wall) due to radial-aware symmetry modeling.
• Ablation findings –
  • FCRA: Reduces false positives, enforces single-connected myocardial contours.
  • ISAI-driven SISD: Improves cross-angular view consistency, enhancing PIA.
  • Prompt-guided modulation: Steers segmentation toward clinically plausible shapes, improving downstream 3D reconstruction via GHD.
• Generalization – Structured integration of anatomical priors improves robustness under motion artifacts and acoustic dropout.

---

📊 Results: Biomarker Estimation

We estimate Ejection Fraction (EF) and Global Longitudinal Strain (GLS) as clinical biomarkers
(calculation details 👉 FeEcho4D). Ablation experiments compare three approaches:

(a) 2D segmentation → Simpson’s 1/3 rule (clinical standard)
(b) Volumetric 3D segmentation
(c) Radial segmentation → GHD-based 3D reconstruction

📌 Key Findings: EF Estimation

• Correlation with ground truth –

  • SCOPE-Net + GHD: 0.888
  • 2D segmentation+Simpson’s: 0.644
  • Volumetric 3D segmentation: 0.731

• Limitations of 2D segmentation+Simpson’s method – In amyloidosis cases, 2D-based EF produced implausible results (including negative EF) due to asymmetric wall thickening and underestimation of EDV from shape assumptions.

• 3D mesh advantage – GHD-based 3D reconstruction accurately quantified ventricular volumes and improved EF precision across the cohort, even in complex morphologies.

• Why SCOPE-Net works better – Symmetry-aware anatomical priors yield contiguous, topologically valid MYO contours, reducing PIA and enabling dense, coherent meshes.

• Robust functional estimation – High-quality meshes preserve spatial integrity across time, supporting accurate EF calculation even in artifact-prone or anatomically ambiguous cases.

---

Citation

If you find this work helpful, please cite:

@article{hasan2025feecho4d,
  title={4D Reconstruction of Fetal Left Ventricle from Echocardiography via 2.5D Radial Segmentation and Graph-Fourier Reconstruction},
  author={XXX},
  journal={XXX},
  volume={XXX},
  pages={XXX},
  year={2025},
  doi={XXX}
}

---

🙏 Acknowledgements

• 👏 We thank all co-authors for their contributions to this work, particularly in model development, dataset construction, and clinical validation.
• 👏 Special thanks to Kepler University Hospital for their support in data acquisition and expert annotations.
• 👏 And to Imperial College London and Dalian University of Technology for providing research infrastructure and technical guidance.
• [Wecome to Dataset Website] [Wecome to Qifeng's Github] [Wecome to Haziq's Github] [Wecome to Yihao's Github]