EUReCA: End-to-End Unified CBCT Reconstruction Across Clinical Acquisitions
Pretrained weights for EUReCA, a single geometry-aware feed-forward network that reconstructs a cone-beam CT volume (256 × 256 × 128 at 2 mm) from 2 to 100 projections on a clinical Varian Halcyon geometry: full or partial arcs, uniform or arbitrary gantry angles, one set of weights per acquisition regime.
- Code: https://github.com/Jiening666/EUReCA
- Project page / paper: https://jiening666.github.io/cbct_recon/
- Demo projections: https://huggingface.co/datasets/jzhu35/EUReCA-demo
Models
All five models share one architecture (210 M parameters, fp32 safetensors, 0.84 GB each) and differ only in the view-sampling regime they were trained on.
| folder | trained for | views | arc |
|---|---|---|---|
eureca_main | general sparse-view CBCT | 8–100, log-uniform | 211° Halcyon half-scan, random start |
eureca_4view | 4 fixed gantry angles | 4 | 0 / 45 / 90 / 135° |
eureca_2ortho | 2 orthogonal views | 2 | 0 / 90° |
eureca_arc90 | limited arc | 15–45 | 90° |
eureca_arc120 | limited arc | 15–45 | 120° |
Each folder contains latest_net_<module>.safetensors for the 17 sub-networks
(encoder, ModLN, view aggregators, 3-D refinement transformer, FDK encoder and
gates, decoder), train_opt.ini with the complete training configuration and
manifest.json with per-module parameter counts and sha256 checksums.
Usage
git clone https://github.com/Jiening666/EUReCA.git && cd EUReCA
pip install -r requirements.txt
python download_weights.py --model eureca_main --demo
bash demo/run_demo.sh eureca_main 20
Or fetch a folder directly:
from huggingface_hub import snapshot_download
snapshot_download("jzhu35/EUReCA", allow_patterns=["eureca_main/*"], local_dir="weights")
Input: (N, 384, 768) line-integral projections resampled to 0.7273 mm at the
isocenter plane (SAD 1000 mm) plus gantry angles in degrees. Output: volume in
[0, 1] mapping to HU −1000 … 1000. See the GitHub README for the exact
conventions and for Varian XIM / iTools preprocessing.
Training data
Trained on digitally reconstructed radiographs of public CT datasets (LUNA16, AMOS22, MELA 2022, RibFrac, TotalSegmentator, AbdomenAtlas 3.0 Mini, RSNA 2023 Abdominal Trauma) and institutional planning CTs, validated on real Halcyon HyperSight projections. No patient data are distributed.
Intended use and limitations
Research use in CBCT reconstruction. The models are tuned to the Halcyon on-board-imager geometry (SAD 1000 mm, SID 1540 mm, 384 × 768 detector at 0.7273 mm isocenter pitch, 2 mm voxels); other geometries need resampling to this detector grid and may degrade. Not a medical device; not for clinical decision-making.
License
CC BY-NC 4.0. The 2D encoder is fine-tuned from Meta AI's ConvNeXt V2
(convnextv2_nano.fcmae_ft_in22k_in1k, CC BY-NC 4.0), so the derived weights
inherit the non-commercial restriction. The source code is Apache-2.0.
Citation
@article{zhu2026eureca,
title = {EUReCA: End-to-End Unified CBCT Reconstruction Across Clinical Acquisitions},
author = {Zhu, Jiening and Zhang, Chengzhu and Fan, Jason and Kuo, LiCheng and Cai, Weixing
and He, Xiuxiu and Cervino, Laura and Moran, Jean and Li, Xiang and Li, Tianfang and Fu, Yabo},
journal = {IEEE Transactions on Medical Imaging (under review)},
year = {2026}
}