VecHeart

By Yihong Chen, Pascal Fua.

This repo is an official implementation of the paper "VecHeart: Holistic Four-Chamber Cardiac Anatomy Modeling via Hybrid VecSets" accepted by MICCAI 2026.

Introduction

Accurate cardiac anatomy modeling requires handling the intricate interrelations among the heart's structures, and VecHeart is a unified framework for holistic reconstruction and generation of four-chamber cardiac anatomy.

Installation

Please refer to INSTALL.md for installation instructions.

Data

We provide the curated 4-chamber cardiac mesh dataset at Google Drive. All cardiac meshes are registered to a common reference frame. Please place (or symlink) the downloaded data under the data/ folder, then run the following preprocessing script in order to generate the spatial sample points and signed distance fields:

python tools/preprocess/sdf_generation.py      # generate the SDF data
python tools/preprocess/generate_slice_pts.py  # generate the multi-view slice points
python tools/preprocess/volume_generation.py   # (optional, need GPU) generate the volume data. It is only used for volume IoU evaluation

After preprocessing, the data/ folder will be structured as follows:

data/
└── VHMesh/
    ├── VHMesh_0001/                    # one folder per sample
    │   ├── VHMesh_0001_1.obj           # 1 = MYO-LV, 2 = LV, 3 = RV, 4 = LA, 5 = RA
    │   ├── VHMesh_0001_2.obj
    │   ├── VHMesh_0001_3.obj
    │   ├── VHMesh_0001_4.obj
    │   ├── VHMesh_0001_5.obj
    │   ├── VHMesh_0001_sdf.npz         # per-part SDF samples       
    │   ├── VHMesh_0001_slice_pts.npz   # 2D multi-view slice contours         
    │   ├── VHMesh_0001_volume.npz      # voxel label volume, for Volume-IoU   
    │   ├── VHMesh_0001_volume.nii.gz   # same volume as NIfTI, for inspection
    │   └── slice_pts/                  # slice point clouds (visualization only)
    │       ├── 4CH.obj
    │       ├── 3CH.obj
    │       ├── 2CH.obj
    │       └── SAX.obj
    ├── VHMesh_0002/
    │   └── ...
    └── VHMesh_1072/
        └── ...

Note: All meshes are kept at their original (world) size on disk. Before sampling points, however, each mesh is first normalized into the cube $[-1, 1]^3$ by a fixed shift and scale:

$$p_{\text{norm}} = (p_{\text{world}} - \text{shift}) \times \text{scale}, \qquad \text{shift} = (13.5997,\ 9.3158,\ -1.0133), \qquad \text{scale} = \tfrac{1}{100}.$$

Reconstruction is therefore carried out in this normalized space. When reporting metrics in world units (e.g. Chamfer distance in mm²), remember to undo the scale — for instance by passing cham_dis_scale to tools/test.py.

Our dataset is curated from multiple sources [1-7]. If you find these data helpful, please remember to cite them.

1. Zeng, A., Wu, C., Lin, G. et al.: Imagecas: A large-scale dataset and benchmark for coronary artery segmentation based on computed tomography angiography images. CMIG (2023)
2. Hansen, B., Pedersen, J., Kofoed, K.F. et al.: A public cardiac ct dataset featuring the left atrial appendage. In: STACOM (2025)
3. Zhuang, X., Li, L., Payer, C. et al.: Evaluation of algorithms for multi-modality whole heart segmentation: an open-access grand challenge. MIA (2019)
4. Metz, C., Schaap, M., Weustink, A. et al.: Coronary centerline extraction from ct coronary angiography images using a minimum cost path approach. Medical physics (2009)
5. chaap, M., Metz, C.T., van Walsum, T. et al.: Standardized evaluation methodology and reference database for evaluating coronary artery centerline extraction algorithms. MIA (2009)
6. Kirişli, H., Schaap, M., Metz, C. et al.: Standardized evaluation framework for evaluating coronary artery stenosis detection, stenosis quantification and lumen segmentation algorithms in computed tomography angiography. MIA (2013)
7. Tobon-Gomez, C., Geers, A.J., Peters, J. et al.: Benchmark for algorithms segmenting the left atrium from 3d ct and mri datasets. TMI (2015)

Get Started

Train

# python tools/train.py <config>
python tools/train.py vecheart.py

Test / Evaluate

# python tools/test.py <config> <checkpoint> --eval
python tools/test.py vecheart.py work_dirs/vecheart/latest.pth --eval

Inference / Visualization

python tools/test.py <config> <checkpoint> --show --show-dir <output_dir>

Configs live under configs/. The provided official configs are explained below:

• VecHeartACM is our full setting.
• VecHeart is trained without ACM, so it can't generalize to missing-part input.
• VecHeartSliceClean is trained on complete synthesized 2CH, 4CH, SAX slice points.
• VecHeartSliceViewDrop randomly drops some views during training, so it can perform inference with an arbitrary combination of views.
• VecHeartSliceViewDropMotionCorrupt additionally introduces the effect of motion corruption on top of the former.

The five configs above are thin leaves that compose shared building blocks from configs/base/ (models/, datasets/, default_runtime.py) via the _base_ inheritance mechanism and override only what differs. If you want to modify something, simply add new models, datasets, metrics and register them, then create a new config — inherit the relevant _base_ pieces and override the keys you need — and you are ready to go.

Note: For the slice-based models, train VecHeartACM model first.

Note: The results you obtained may differ from the paper a bit. That's expected as we change the dataset processing flow for public release.

Acknowledgement

Thanks to these great repositories: mmcv, VecSetX, sdf_gen.

Citation

If you find this work useful in your research, please consider citing:

@article{chen2025tetheart,
  title   = {VecHeart: Holistic Four-Chamber Cardiac Anatomy Modeling via Hybrid VecSets},
  author  = {Chen, Yihong and Fua, Pascal},
  journal = {MICCAI},
  year    = {2026}
}