CaiYuanhao/X-Gaussian

Radiative Gaussian Splatting for Efficient X-ray Novel View Synthesis

ECCV 2024

5 CT scenes + 3 real scans · X-ray novel view synthesis · Direct 3D CT reconstruction

Paper · Code · Models · Data

Point Cloud Visualization

Training Process Visualization

GIF demonstrations are from the official X-Gaussian repository. The five scenes available in this release are listed below.

Dataset Description

This repository hosts the five processed CT scenes used in X-Gaussian, introduced in Radiative Gaussian Splatting for Efficient X-ray Novel View Synthesis (ECCV 2024): Chest, Abdomen, Foot, Head, and Pancreas.

Each scene contains a reference 3D CT volume, simulated cone-beam X-ray projections, view angles, and scanner geometry. The data support sparse-view reconstruction and held-out X-ray projection rendering. The current X-Gaussian code also supports direct 3D CT reconstruction from the same Gaussian field.

Performance

X-ray Novel View Synthesis

Results measured on the hosted joint 2D/3D checkpoints, averaged over 50 held-out views per scene:

Scene	2D PSNR (dB)	2D SSIM
Chest	47.21	0.9937
Abdomen	48.16	0.9968
Foot	44.40	0.9869
Head	51.66	0.9979
Pancreas	47.65	0.9923

Each predicted and reference projection is normalized by its own maximum.

Original paper comparisons

The figures reproduce the original paper results; the table above evaluates the hosted joint 2D/3D checkpoints.

Direct 3D CT Reconstruction

Results measured on the five released models after 30,000 training steps and 1,000 refinement steps, without downsampling:

Scene	CT resolution	3D PSNR (dB)	3D SSIM
Chest	128 × 128 × 128	31.94	0.8895
Abdomen	512 × 512 × 463	33.33	0.8917
Foot	256 × 256 × 256	31.88	0.8795
Head	256 × 256 × 128	40.21	0.9732
Pancreas	512 × 512 × 240	22.22	0.8750

3D metrics use a fixed range of 1 on the original voxel grid, without clipping or intensity correction. SSIM averages nonempty reference slices along each axis, then across axes. CT ground truth is used only for evaluation.

Real Data

Three real FIPS cone-beam scans are available in real/: Pine, Seashell, and Walnut. Each scene contains 50 training projections, 100 held-out measured projections, 560 × 560 detector images, and a 256³ full-view FDK pseudo-reference. The official processed projections, scanner geometry, split, and reference volume are preserved without additional downsampling. The provided initialization is computed by X-Gaussian from only the 50 training views; the pseudo-reference is used only for evaluation.

See real-data contents, original sources, and CC BY 4.0 attribution. The 3D reference is an algorithmic reconstruction, not measured CT ground truth.

Dataset Contents

One Python pickle file is provided per original CT scene. The real scans use the folder format described above. Dimensions below are the original stored array dimensions; no downsampling is applied to this release.

File	CT volume shape	Projection shape per view
chest_50.pickle	128 × 128 × 128	256 × 256
abdomen_50.pickle	512 × 512 × 463	1024 × 1024
foot_50.pickle	256 × 256 × 256	512 × 512
head_50.pickle	256 × 256 × 128	512 × 512
pancreas_50.pickle	512 × 512 × 240	1024 × 1024

Each file contains 50 training projections and 50 held-out projections. The held-out split is named val; these are different views of the same scene, rather than separate training and validation scenes. The uploaded files specify cone-beam geometry, a configured scan range of 180°, and noise: 0.

Data Format

Field	Description
image	Reference CT volume with shape (N_x, N_y, N_z), used for reconstruction evaluation.
train.projections, val.projections	Projection arrays with shape (50, H, W).
train.angles, val.angles	Corresponding view angles in radians.
numTrain, numVal	Numbers of training and held-out views.
DSD, DSO	Source-to-detector and source-to-origin distances.
nDetector, dDetector	Detector pixel counts and pixel sizes.
nVoxel, dVoxel	CT voxel counts and voxel sizes.
offOrigin, offDetector	Volume and detector offsets.
mode, accuracy, filter	Acquisition and reconstruction settings.

Physical lengths in the stored scanner metadata are in millimeters. The official loader handles unit conversion and scene normalization. CT ground truth is not required by the current X-Gaussian training workflow.

Usage

Please refer to the X-Gaussian repository for installation, data preparation, training, and testing.

Data Sources and Processing

These scenes use the processed CT data format shared with SAX-NeRF. X-ray projections are synthesized from CT volumes using the TIGRE-based data generation pipeline. See the SAX-NeRF dataset card and data generation code for source attribution and preprocessing details. Acquisition settings for this release should be read from the individual files as summarized above.

Citation

If you use these data, please cite X-Gaussian and SAX-NeRF:

@inproceedings{x_gaussian,
  title={Radiative gaussian splatting for efficient x-ray novel view synthesis},
  author={Yuanhao Cai and Yixun Liang and Jiahao Wang and Angtian Wang and Yulun Zhang and Xiaokang Yang and Zongwei Zhou and Alan Yuille},
  booktitle={ECCV},
  year={2024}
}

@inproceedings{sax_nerf,
  title={Structure-Aware Sparse-View X-ray 3D Reconstruction},
  author={Yuanhao Cai and Jiahao Wang and Alan Yuille and Zongwei Zhou and Angtian Wang},
  booktitle={CVPR},
  year={2024}
}

License

The original five CT datasets retain the Apache-2.0 license. The FIPS real data and derived initialization under real/ are CC BY 4.0; see source attribution and processing details.