lyqun/SPARC-reconstruction

SPARC — reconstruction models

Sixteen SPARC models that reconstruct a 256³ CT volume from four or eight X-ray projections, one per dataset and view count. They are the models behind every reconstruction result in the manuscript A foundation model recovers three-dimensional anatomy and clinical findings from sparse X-ray projections (2026). Each was adapted from the pretrained backbone in lyqun/SPARC.

Code, installation and data preparation: HORIZONHealthcare/SPARC.

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.

Performance

Mean over each test set, with patient-level bootstrap 95% confidence intervals (10,000 resamples). PSNR and SSIM (× 100) are computed over [−1,024, 1,024] HU. The last column is SPARC's paired lead in PSNR over DeepSparse, the strongest of five competing methods retrained under the same protocol in the paper; its interval excluded zero in every setting.

Dataset	Projections	Test volumes	PSNR (dB)	SSIM (× 100)	Lead over DeepSparse (dB)
CT-RATE	4	602	29.41 (29.25–29.56)	87.15 (86.76–87.53)	+2.25 (+2.16 to +2.35)
CT-RATE	8	602	31.39 (31.25–31.54)	89.18 (88.85–89.50)	+2.07 (+2.01 to +2.14)
TotalSegmentator	4	79	29.37 (28.98–29.76)	87.46 (86.74–88.17)	+2.74 (+2.59 to +2.89)
TotalSegmentator	8	79	31.37 (31.03–31.72)	90.46 (89.86–91.05)	+2.62 (+2.50 to +2.73)
Medical Segmentation Decathlon	4	209	30.27 (29.98–30.55)	86.98 (86.51–87.42)	+1.84 (+1.76 to +1.92)
Medical Segmentation Decathlon	8	209	32.74 (32.43–33.03)	92.09 (91.66–92.50)	+1.88 (+1.81 to +1.96)
AbdomenCT-1K	4	158	29.74 (29.44–30.04)	88.45 (87.90–88.99)	+1.70 (+1.62 to +1.79)
AbdomenCT-1K	8	158	32.23 (31.91–32.53)	90.17 (89.82–90.50)	+1.58 (+1.49 to +1.67)
AMOS (CT)	4	75	29.65 (29.37–29.91)	90.18 (89.78–90.57)	+1.65 (+1.49 to +1.81)
AMOS (CT)	8	75	31.67 (31.40–31.93)	92.18 (91.84–92.50)	+0.83 (+0.72 to +0.94)
CQ500	4	71	30.86 (30.42–31.24)	95.59 (94.65–96.21)	+1.37 (+1.19 to +1.56)
CQ500	8	71	33.06 (32.66–33.40)	96.89 (96.14–97.37)	+1.10 (+0.94 to +1.30)
ToothFairy3	4	77	30.82 (30.47–31.16)	95.46 (95.09–95.80)	+1.46 (+1.34 to +1.58)
ToothFairy3	8	77	33.30 (32.94–33.66)	96.64 (96.35–96.90)	+1.70 (+1.57 to +1.82)
VerSe	4	50	27.61 (26.75–28.50)	81.43 (79.33–83.52)	+2.07 (+1.88 to +2.27)
VerSe	8	50	29.48 (28.73–30.25)	86.23 (84.48–87.90)	+2.21 (+1.98 to +2.46)

Thirteen test volumes that repeat a training scan, or come from a subject with a training volume, were removed from the test sets before evaluation (3 MSD, 1 AbdomenCT-1K, 3 ToothFairy3, 6 VerSe); they are listed in splits/excluded_test_volumes.csv.

Files

Each checkpoint (544 MB) holds the weights of the 2D projection encoder, the 3D feature volume and the point decoder (136.1 M parameters, float32), the training configuration (config), and the epoch and validation PSNR of the checkpoint, which was selected on validation PSNR. It loads with torch.load(..., weights_only=True).

File	Epoch	Validation PSNR (dB)	SHA-256
ctrate_v4.pth	380	29.23	b6728802b21dc2cd397b50189417f2db453bbfd39d4edfb81d56baef8880ad51
ctrate_v8.pth	400	31.14	c965b28bb87d6db002fe0637a299dc8bf991b0f1b27d0d3ebaeb9b1b1649a917
totalsegmentator_v4.pth	380	30.03	4377e98910f3e7d37d46503cc088dddd2898b291afd974a17def9e3d6b219a87
totalsegmentator_v8.pth	400	31.90	17cf829df910a697ca37975719c497f5c69f23b5b2e770edf3c1cde535ceeaff
msd_v4.pth	180	29.90	64da7d6e8e3a46059dcd5ad9d0cf4c6a9e4960d7b9bd3b27d18808fb13cd54f6
msd_v8.pth	240	32.28	72eb874fbcb5257fb098dabd27831796ef1d17de9b936c66af801697e686aef0
abdomenct1k_v4.pth	240	29.58	830b79fff54b3a98b02003ab867804972e6963602c8c066c16318102ddfba300
abdomenct1k_v8.pth	240	32.11	6b2c17fc3844565554ac2da90056fd9d6a9aeb389732d95b88c72a397040dfbe
amos_v4.pth	360	29.65	1ee20863aac055871ab731c02ce6dd1ef775b89f8ee392f5281cdaec7a6f2277
amos_v8.pth	400	31.68	d60d040ec5af2896f67db1bb13b95a1ef27733cd303009d928566e23e3cf2b8f
cq500_v4.pth	360	30.32	4ef547394eeb70a50881fa4784b29cde28b0dafac68995dcbcbb6df7d6c3eacc
cq500_v8.pth	360	32.76	62e6d30249c83dbff7f48e16f3e37ff61b090e79468bbcfd898015c01ad41c98
toothfairy3_v4.pth	360	30.81	133d21ac82593aa3e95ecbb28ef3a5d42c0d954ff6d90da1c461b8f8cdd1cecc
toothfairy3_v8.pth	400	33.30	8f7ecb8db40f87aed2e8ae4b7a79f2b6dc81e2f3d4a3804ea5091b69acaba634
verse_v4.pth	80	27.81	3d7bd5d1017733f6533e0ed3838e42e875e62b8d94553f616dafd75748d89175
verse_v8.pth	80	29.63	d4bba3d07ff0d251abdec811f2f3b6b5453b16a38a08bcd25b8ee83566204c05

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
hf download lyqun/SPARC-reconstruction cq500_v8.pth --local-dir weights

export DATA_ROOT=/path/to/data OUTPUT_ROOT=/path/to/outputs PYTHONPATH=.
cp -r splits "$DATA_ROOT/"
python finetune_foundation_recon.py --config configs/cq500_v8.yaml \
    --init_ckpt weights/sparc_stage2_backbone.pth --eval_ckpt weights/cq500_v8.pth \
    --metrics_manifest "$DATA_ROOT/splits/cq500_test.csv" --metrics_csv cq500_v8_metrics.csv

The backbone file supplies the architecture; the reconstruction file supplies the weights. Use --export_dir <folder> instead of --metrics_csv to save reconstructed volumes in Hounsfield units. Each model expects its own view count and the acquisition geometry of its dataset, which the config sets:

Files	Config	Test split
ctrate_v{4,8}.pth	configs/ctrate_v{4,8}.yaml	splits/ctrate_test.csv
totalsegmentator_v{4,8}.pth	configs/totalsegmentator_v{4,8}.yaml	splits/totalsegmentator_test.csv
msd_v{4,8}.pth	configs/msd_v{4,8}.yaml	splits/msd_test.csv
abdomenct1k_v{4,8}.pth	configs/abdomenct1k_v{4,8}.yaml	splits/abdomenct1k_test.csv
amos_v{4,8}.pth	configs/amos_v{4,8}.yaml	splits/amos_test.csv
cq500_v{4,8}.pth	configs/cq500_v{4,8}.yaml	splits/cq500_test.csv
toothfairy3_v{4,8}.pth	configs/toothfairy3_v{4,8}.yaml	splits/toothfairy3_test.csv
verse_v{4,8}.pth	configs/verse_v{4,8}.yaml	splits/verse_test.csv

The volumes must first be converted as described in preprocessing/README.md.

With the released code, weights and split files, all sixteen models reproduce the per-volume results used in the paper.

Training data

Each model was finetuned on the training split of one dataset, starting from the pretrained backbone, for at most 400 epochs with early stopping on validation PSNR. The splits are in splits/.

Dataset	Anatomy	Source	Licence
CT-RATE	Chest	source	CC BY-NC-SA 4.0
TotalSegmentator v2.0.1	Multiple regions	source	CC BY 4.0
Medical Segmentation Decathlon	Abdomen and chest	source	CC BY-SA 4.0
AbdomenCT-1K	Abdomen	source	see source
AMOS (CT)	Abdomen	source	CC BY 4.0
CQ500	Head (unseen in pretraining)	source; we used the Kaggle mirror	CC BY-NC-SA 4.0
ToothFairy3	Dental CBCT (unseen in pretraining)	source	CC BY-NC-SA 4.0
VerSe	Spine (unseen in pretraining)	source; we used the copy in CADS (0010_verse)	CC BY-SA 4.0

Limitations

• 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.
• Each model assumes the acquisition geometry of its dataset and view count; a miscalibrated geometry degrades it sharply (in the paper's simulations, up to 13.3 dB in PSNR for a 5% source-to-detector distance error or a 90° arc).
• A model trained on one dataset has not been validated on scanners, populations or anatomies outside it.
• For research use only. They are not medical devices.

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 each model was 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 the dataset each model was trained on.