How to Choose a CT Reconstruction Dataset
Last updated:
Choose a CT reconstruction dataset by asking four questions in order: does it ship raw projections, does its geometry match yours (2D fan-beam, helical or cone-beam), is it measured or simulated, and does its licence allow your use. Of the 21 datasets with a page here, 13 include projection data, which is what you need to test a reconstruction method rather than an image post-processor.
The table at the end compares all of them; each name links to a page with the full record and the papers that use it.
1. Projections or images only
A reconstruction method needs sinograms. The directory records what each dataset contains: of the dataset pages here, 8 ship projections and images, 5 ship projections only, and 1 are image collections; the rest do not say. Image-only collections such as ToothFairy2 are still useful for pretraining and for downstream evaluation.
2. Geometry
5 of the datasets here are 2D fan-beam, which is quick to iterate on; 9 are cone-beam, which is what cone-beam CT methods need. A method developed on 2D slices does not automatically carry over to a 3D cone-beam geometry.
3. Measured or simulated
Simulated data give exact ground truth and any protocol you like: LoDoPaB-CT is simulated from LIDC/IDRI, and the AAPM DL-Sparse-View CT data use noiseless phantoms. Measured data show whether a method survives real acquisitions: 2DeteCT has real experimental slices at three dose levels, and the Helsinki Tomography Challenge 2022 data are real limited-angle sinograms.
4. Licence and access
7 of the datasets here carry a NonCommercial (NC) licence, and some parts need an access request: the head scans in LDCT-and-Projection-data need an NIH controlled-access request. The licence column describes each dataset, not this list; check the official page before you redistribute anything.
All datasets compared
Every dataset with a page here, projection data first. Benchmarks built on these datasets are on the datasets and benchmarks page.
| Name | Data | Modality | Licence |
|---|---|---|---|
| LoDoPaB-CT | Projection + image | 2D fan-beam | CC BY 4.0 |
| LDCT-and-Projection-data | Projection + image | helical CT | CC BY 4.0 (head scans NIH controlled access) |
| 2DeteCT | Projection + image | 2D fan-beam | CC BY 4.0 |
| SPARE 4D-CBCT | Projection + image | cone-beam, 4D | research use |
| AAPM DL-Sparse-View CT data | Projection + image | 2D fan-beam (simulated) | CC BY 4.0 |
| COBRA2026 | Projection + image | cone-beam | CC BY-NC 4.0 |
| SparseBeads | Projection + image | micro-CT | CC BY 4.0 |
| FleX-ray Walnut CBCT Collection | Projection + image | cone-beam micro-CT | CC BY 4.0 |
| Walnut-CBCT | Projection only | cone-beam | CC BY 4.0 |
| Helsinki Tomography Challenge 2022 | Projection only | 2D fan-beam | CC BY 4.0 |
| FIPS tomography datasets | Projection only | fan-beam / cone-beam | CC BY 4.0 |
| Walnut-21 (two dose levels) | Projection only | cone-beam | CC BY 4.0 |
| Carved cheese (FIPS) | Projection only | 2D fan-beam | CC BY 4.0 |
| ToothFairy2 | Image only | cone-beam, dental | CC BY 4.0 |
| CBCT Head & Neck (TCIA) | Not stated | cone-beam | CC BY-NC 3.0 |
| LIDC-IDRI | Not stated | helical CT | CC BY 3.0 |
| STOIC 2021 | Not stated | CT | CC BY-NC 3.0 |
| MosMedData | Not stated | CT | CC BY-NC-ND 3.0 |
| CTPelvic1K | Not stated | contrast CT | CC BY-NC 4.0 |
| MORE | Not stated | CT | CC BY-NC 4.0 |
| LoDoInd | Not stated | industrial cone-beam CT | CC BY-NC 4.0 |
Related pages
- CT Reconstruction Datasets and Benchmarks
- Sparse-View CT Reconstruction: Methods, Datasets and Where to Start
- Low-Dose CT Denoising: Benchmarks, Datasets and Toolkits
Frequently asked questions
Which CT datasets include raw projection data?
Among the datasets with a page here: LoDoPaB-CT, LDCT-and-Projection-data, 2DeteCT, Walnut-CBCT, Helsinki Tomography Challenge 2022, SPARE 4D-CBCT, AAPM DL-Sparse-View CT data, FIPS tomography datasets, COBRA2026, SparseBeads, Walnut-21 (two dose levels), FleX-ray Walnut CBCT Collection, Carved cheese (FIPS).
Which dataset should I use for limited-angle CT?
The Helsinki Tomography Challenge 2022 data are real limited-angle sinograms, and the accompanying challenge defines the standard limited-angle protocol. For sparse-view CT, see the sparse-view guide.
Can I use these datasets commercially?
Not all of them: 7 of the datasets here carry a NonCommercial (NC) licence. The licence shown on each page is the dataset's own, not this list's; check the official page for the current terms.
What is the difference between a dataset and a benchmark here?
A dataset entry is a collection of scans. A benchmark entry is a challenge, benchmark or standard evaluation protocol, often built on a dataset: the LoDoPaB-CT benchmark, for example, evaluates on the LoDoPaB-CT split. Both are listed on the datasets and benchmarks page.