This repository contains the official implementation of DeepSparse: A Foundation Model for Sparse-View CBCT Reconstruction , accepted in IEEE Transactions on Medical Imaging (TMI), 2026.

For any questions regarding this repository, please contact jchenhu@connect.ust.hk.

• The code for pretraining, finetuning, and evaluation is released.
• Pretrained and finetuned model weights are released on HuggingFace.

• Foundation model for sparse-view CBCT: DeepSparse is pretrained on a large-scale abdominal CT dataset (AtlasAbdomenMini, ~7k volumes) and can be finetuned to diverse anatomical targets with few labelled examples.
• Implicit neural representation with vector quantization: A codebook-based encoder-decoder reconstructs 3D volumes from sparse 2D X-ray projections, enabling memory-efficient inference at arbitrary resolution.
• Two-stage finetuning: Stage 1 adapts the pretrained backbone to a target anatomy; Stage 2 refines the fine detail decoder with a frozen encoder, enabling high-quality reconstruction with as few as 6 views.
• Generalizes across anatomies: Evaluated on abdomen (PANORAMA), pelvis (PENGWIN), lung (LUNA16), and tooth (ToothFairy) with consistent state-of-the-art performance.

git clone https://github.com/xmed-lab/DeepSparse.git
cd DeepSparse
conda create -n deepsparse python=3.9
conda activate deepsparse
# PyTorch 1.13.1 + CUDA 11.6
conda install pytorch==1.13.1 torchvision==0.14.1 torchaudio==0.13.1 pytorch-cuda=11.6 -c pytorch -c nvidia
# Other dependencies
pip install SimpleITK scipy numpy einops tqdm matplotlib pytorch-lightning timm
# TIGRE (cone-beam CT projector) — follow the instructions at:
# https://github.com/CERN/TIGRE

All dataset preprocessing scripts are in data/. See data/README.md for the full directory structure and shared utilities.

Each dataset has its own subfolder with a README.md:

Tip: You can skip training entirely and download pretrained weights for direct inference — see Download Checkpoints below.

bash scripts/pretrain.sh

Trains the foundation model on AtlasAbdomenMini for 800 epochs. The checkpoint used for finetuning is saved at logs/pretrain/ep_700.pth.

bash scripts/finetune_s1.sh

Adapts the pretrained backbone to each downstream anatomy (abdomen, pelvis, lung, tooth) across three sparsity regimes (6, 8, 10 views). Checkpoints are saved under logs/{dataset}+{n_view}v+s1/.

bash scripts/finetune_s2.sh

Refines with a frozen encoder to sharpen reconstruction quality. Loads Stage 1 checkpoints and saves results under logs/{dataset}+{n_view}v+s2/.

bash scripts/evaluate_s2.sh

Runs evaluation on all datasets and view counts using Stage 2 checkpoints at logs/{dataset}+{n_view}v+s2/ep_400.pth.

Pretrained and finetuned checkpoints are hosted on HuggingFace and can be downloaded for direct inference:

HuggingFace — HajihajihaJimmy/DeepSparse

Place the downloaded logs/ folder in the project root. See logs/README.md for a description of each checkpoint.

To download with huggingface-cli:

pip install huggingface-hub hf-transfer
HF_HUB_ENABLE_HF_TRANSFER=1 huggingface-cli download HajihajihaJimmy/DeepSparse --local-dir ./logs

If you find this work helpful, please cite:

@ARTICLE{11436116,
 author={Lin, Yiqun and Chen, Jixiang and Wang, Hualiang and Yang, Jiewen and Guo, Jiarong and Zhang, Yi and Li, Xiaomeng},
 journal={IEEE Transactions on Medical Imaging},
 title={DeepSparse: A Foundation Model for Sparse-View CBCT Reconstruction},
 year={2026},
 volume={},
 number={},
 pages={1-1},
 doi={10.1109/TMI.2026.3674948}
}

We thank the challenge organizers of PANORAMA, PENGWIN, LUNA16, and ToothFairy for making their datasets publicly available.