Diffusion Models
DM4CT (Real-World Synchrotron CT)
The DM4CT real-world synchrotron CT data sets accompany the manuscript “DM4CT: Benchmarking Diffusion Models for Computed Tomography Reconstruction” [B40]. They contain synchrotron X-ray CT scans of two rock samples. For each sample, raw projection data and reconstructed slices are provided. The reconstructions include flat-field correction (using median dark and flat fields), log transformation, and ring artifact reduction.
DM4CT was designed as a benchmark for AI-based CT reconstruction, with a focus on diffusion models used as learned priors/solvers for inverse problems. The benchmark evaluates diffusion-based reconstruction methods under realistic conditions (e.g., sparse-view and noisy measurements) and compares them against strong non-AI baselines (e.g., filtered backprojection and model-based iterative reconstruction). The real-world synchrotron data sets provided here enable evaluation of AI reconstruction methods on experimental data using a consistent forward model and evaluation protocol described in the DM4CT paper.
If you use these data sets, please cite the DM4CT paper at [B40].
Instrument |
Elettra Sincrotrone Trieste |
Energy |
24 keV |
Filter |
None |
Scan Range |
180 degree |
Number of Projections |
1200 |
Angular Step |
0.15 degree |
White Fields |
20 |
Dark Fields |
20 |
Exposure Time |
4 s |
Sample Detector Distance |
150 mm |
Pixel Size |
9 µm |
To reproduce the preprocessing/postprocessing used for the provided reconstructions (flat-field correction, log transformation, rotation, ring artifact reduction), refer to the DM4CT code repository:
https://github.com/DM4CT/DM4CT
Zenodo record hosting the original release:
https://zenodo.org/records/15420527
Tomo ID |
Sample |
Contents |
Rotation axis location |
Projection Preview |
Reconstruction Preview |
F3_1 (rock sample) |
projections + recon slices |
-61 |
|
|
|
F3_2 (rock sample) |
projections + recon slices |
-62 |
|
|
Notes
Each scan provides 1200 projections plus 20 dark fields and 20 flat fields.
The provided reconstructions include flat-field correction (median dark/flat), log transformation, and ring artifact reduction.
The DM4CT repository includes scripts that reproduce the cropped reconstructions used in the accompanying paper.