HeartCompv1

HeartCompv1: A Large-Scale Multi-Class Cardiac Point Cloud Completion Dataset

HeartCompv1 is a large-scale dataset for 3D cardiac shape reconstruction and point cloud completion, containing 17,000 four-chamber cardiac samples. The dataset is designed to support the development and evaluation of geometric deep learning methods for reconstructing complete cardiac anatomy from sparse observations.

Each sample consists of six anatomically and functionally distinct cardiac substructures:

• Left Ventricular (LV) Endocardium
• LV Epicardium
• Right Ventricular (RV) Endocardium
• RV Epicardium
• Left Atrium (LA)
• Right Atrium (RA)

HeartCompv1 provides paired sparse point clouds and corresponding dense ground-truth point clouds, enabling supervised learning and quantitative evaluation of cardiac reconstruction methods.

Dataset Composition

The dataset contains a total of 17,000 samples organized into training, validation, and testing subsets:

• Training set: 10,000 samples with varying degrees of simulated spatial misalignment to improve model robustness.
• Validation set: 1,000 samples with varying degrees of simulated spatial misalignment for model selection and hyperparameter tuning.
• Test sets: 6 independent test sets, each containing 1,000 samples and representing different levels of reconstruction difficulty and spatial misalignment.

The inclusion of multiple test sets allows comprehensive evaluation of model performance under increasingly challenging reconstruction scenarios.

Data Contents

Each sample includes:

• Sparse cardiac point cloud
• Dense ground-truth cardiac point cloud

Notes

HeartCompv1 is intended as a benchmark dataset for developing and evaluating methods that reconstruct complete four-chamber cardiac anatomy from sparse observations. The dataset includes varying levels of spatial misalignment to simulate realistic acquisition and registration challenges and to facilitate robustness assessment across different reconstruction conditions.

Citation

If you use HeartCompv1 in your research, please cite:

Ma, Z., & Banerjee, A. (2026, September). HeartFormer: Semantic-Aware Dual-Structure Transformers for 3D Four-Chamber Cardiac Point Cloud Reconstruction. In International Conference on Medical Image Computing and Computer-Assisted Intervention. Cham: Springer Nature Switzerland.
Ma, Z., & Banerjee, A. (2025). HeartFormer: Semantic-Aware Dual-Structure Transformers for 3D Four-Chamber Cardiac Point Cloud Reconstruction. arXiv preprint arXiv:2512.00264.