SPARC — pretrained backbone

SPARC is a foundation model that reconstructs a three-dimensional CT volume from sparse X-ray projections. This repository holds the pretrained weights from which every SPARC reconstruction model is adapted. They accompany the manuscript A foundation model recovers three-dimensional anatomy and clinical findings from sparse X-ray projections (2026).

Code, installation and data preparation: HORIZONHealthcare/SPARC. The sixteen per-dataset reconstruction models are in lyqun/SPARC-reconstruction.

Access. Log in, fill in the short form on this page, and access is granted straight away. Then log in on your machine with hf auth login before downloading.

Files

File	Contents	Use
sparc_stage2_backbone.pth	Stage-2 model: 2D projection encoder, 3D feature volume and point decoder (136.1 M parameters, 544 MB)	Starting point for reconstruction on a new dataset, or for the released reconstruction models
sparc_stage1_ctmae.pth	Stage-1 3D masked autoencoder on CT, encoder and decoder (340.4 M parameters, 1.36 GB)	Only needed to rerun Stage-2 pretraining, where its encoder is the frozen target

Each file is a plain dictionary with the weights (model, float32), the settings needed to rebuild the model (config) and the training iteration (iter). It loads with torch.load(..., weights_only=True).

File	SHA-256
sparc_stage2_backbone.pth	b9d8605e77458cd1af5024727df33f8a286585432ce1f62ca925c1114c45098d
sparc_stage1_ctmae.pth	f13ddf31cba1fdfabad7708bfb58b303e9c3cacd5269d3aedcd1c92f1b3b8d85

Model

• Stage 1. A 3D masked autoencoder (ViT-L/16 encoder, 0.75 masking ratio, normalised-pixel loss) trained on 256³ CT crops for 200,000 iterations. Its encoder turns a CT volume into 16³ = 4,096 tokens.
• Stage 2. Projections are rendered from CT on the fly with randomised acquisition geometry: 2 to 16 views, uniform, random, limited-angle, bi-planar and clustered angle patterns, and randomised source and detector distances. Each projection passes through a shared ConvNeXt pyramid with a transformer at the coarsest scale, into which the Plücker coordinates of every ray are injected. A 16³ grid of points is projected into all views, sampled, max-fused across views and refined by a 3D transformer into a feature volume of 16³ × 768. A point decoder reads that volume at any query point and predicts Hounsfield units. A second head that predicted the frozen Stage-1 tokens was used during pretraining and is not included, as in the paper. 150,000 iterations on 16 NVIDIA GH200 GPUs, one volume per GPU.

The released backbone reconstructs directly, but the paper's reconstruction results come from per-dataset adaptation of it; for those use lyqun/SPARC-reconstruction.

Training data

Both stages were trained on the official training split of CT-RATE (chest CT; 47,149 volumes from 24,128 scans of 20,000 patients). Every chest evaluation volume in the paper comes from CT-RATE's official validation split, so no pretraining patient appears in any evaluation list. The exact list is splits/ctrate_pretrain.csv.

Usage

git clone https://github.com/HORIZONHealthcare/SPARC.git && cd SPARC
conda create -n sparc python=3.11 -y && conda activate sparc
pip install torch==2.5.1 --index-url https://download.pytorch.org/whl/cu124
pip install -r requirements.txt
hf auth login        # after filling in the form on this page
hf download lyqun/SPARC sparc_stage2_backbone.pth --local-dir weights

# adapt to a dataset, starting from the pretrained backbone
python finetune_foundation_recon.py --config configs/cq500_v8.yaml \
    --init_ckpt weights/sparc_stage2_backbone.pth

To rerun Stage-2 pretraining without Stage 1, download sparc_stage1_ctmae.pth and set target.ctmae_init_path in configs/stage2_backbone.yaml to its path.

Limitations

• SPARC was trained and evaluated on projections rendered from CT (digitally reconstructed radiographs), not on radiographs acquired on an X-ray system. Scatter, detector response and beam hardening on a real system differ from the rendering.
• Reconstruction needs an accurately calibrated acquisition geometry. In the paper's simulations, a 5% error in the source-to-detector distance or a 90° arc cost SPARC up to 13.3 dB in PSNR, far more than photon noise at 10⁴ incident photons per detector pixel (at most 3.4 dB).
• Pretraining used chest CT only.
• For research use only. It is not a medical device.

License

The weights are released under CC BY-NC 4.0, the same licence as the code. Commercial use is not permitted. Use of the weights must also respect the terms of the datasets they were trained on. The models are for research use only and are not medical devices.

Citation

@misc{lin2026sparc,
  title  = {A foundation model recovers three-dimensional anatomy and clinical findings from sparse X-ray projections},
  author = {Lin, Yiqun and Xu, Jiayang and Ju, Lie and Wang, Hualiang and Guo, Jiarong and Yao, Huifeng and Sun, Haoran and Zhou, Yukun},
  year   = {2026},
  note   = {Manuscript}
}

Please also cite CT-RATE: Hamamci I. E. et al. Generalist foundation models from a multimodal dataset for 3D computed tomography. Nature Biomedical Engineering (2026).