Low-Dose CT Denoising: Benchmarks, Datasets and Toolkits
Last updated:
For low-dose CT denoising, evaluate on the Mayo Clinic data from the 2016 AAPM Low-Dose CT Grand Challenge, now released as LDCT-and-Projection-data on TCIA, and on LoDoPaB-CT, the main simulated benchmark. Nearly every low-dose CT paper still reports on the AAPM split. Reconstruct with an open toolkit such as the ASTRA Toolbox, TIGRE, ODL or CIL, and use the LD-CT simulator when you need dose levels a dataset does not have.
The directory lists 69 low-dose CT papers, 20 of them with code. The tables below are generated from those records.
Benchmarks compared
These are the benchmarks and evaluation protocols the low-dose CT topic page lists, with the year each started and what it measures. Each has its own page with the papers in the directory that report on it.
| Name | Year | What it evaluates |
|---|---|---|
| AAPM Low-Dose CT Grand Challenge | 2016 | The challenge that established the quarter-dose / 10% / 5% protocol and the Mayo abdominal cohort. Nearly every low-dose CT paper still reports on this split. |
| LoDoPaB-CT benchmark | 2021 | Grand-challenge style evaluation on the LoDoPaB-CT split, with a published, tuned FBP+U-Net reference baseline so improvements are measurable. |
| SPIE-AAPM-NCI LungX Challenge | 2015 | Lung nodule malignancy classification on low-dose screening CT — the task is diagnostic and benchmarked against radiologists, not nodule detection. |
| Noise2Noise denoising protocol | 2018 | The reference implementation that popularised training a denoiser without clean targets. On CT it became the standard 'no clean ground truth' evaluation protocol — most unsupervised CT denoising… |
Datasets
For projection-domain or dual-domain methods you need raw sinograms; image-domain denoisers can also train on reconstructed slices. The data column says which a dataset ships.
| 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) |
| COBRA2026 | Projection + image | cone-beam | CC BY-NC 4.0 |
| Walnut-21 (two dose levels) | Projection only | cone-beam | CC BY 4.0 |
| 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 |
| LoDoInd | Not stated | industrial cone-beam CT | CC BY-NC 4.0 |
Toolkits
General reconstruction toolkits provide projectors and iterative solvers, PyTorch libraries add learned and unrolled methods, and simulators turn full-dose images into realistic low-dose ones. Licences are those of the toolkits.
| Name | Language | Licence | What it provides |
|---|---|---|---|
| ASTRA Toolbox | C++ / Python / MATLAB | GPL-3.0 | Long-standing research toolbox with a large set of projectors and algorithms; the reference CPU/GPU baseline. |
| TIGRE | MATLAB / Python | BSD-3-Clause | GPU-accelerated (CUDA) toolbox for cone-beam and parallel-beam CT; FDK, SART, SIRT, CGLS and TV variants. |
| ODL | Python | MPL-2.0 | Operator Discretization Library — a functional-analytic framework for inverse problems, with CT geometries built in. |
| Core Imaging Library (CIL) | Python | Apache-2.0 | CCPi framework for variational CT reconstruction with a large catalogue of optimisation algorithms. |
| DeepInverse | Python | BSD-3-Clause | PyTorch library for imaging inverse problems — physics operators, learned denoisers, unrolled networks and benchmarks. |
| LION | Python | GPL-3.0 | Learned Iterative Optimization Networks — unrolled reconstruction with PyTorch, used in CT and PET. |
| LD-CT simulator | MATLAB | See repository | Low-dose CT simulator that converts high-dose images into realistic low-dose counterparts using a validated noise model. |
| Physics-ArX | Python | See repository | Physics-based data-augmentation library that injects realistic CT/CBCT noise, scatter and artefact models into existing images to build paired training data. |
Where to start
The maintainers' start-here picks from the low-dose CT topic page: the standard CNN baseline, a GAN with a perceptual loss, a transformer, the LoDoPaB-CT descriptor and a self-supervised method. More methods that train without clean targets are on the self-supervised topic page.
| Paper | Venue | Year | Why read it |
|---|---|---|---|
| Low-Dose CT with a Residual Encoder-Decoder Convolutional Neural Network · code | TMI | 2017 | The residual encoder-decoder CNN that became the standard image-domain denoising baseline, with code. |
| Low Dose CT Image Denoising Using a Generative Adversarial Network with Wasserstein Distance and Perceptual Loss · code | TMI | 2018 | Introduced Wasserstein GAN training with a perceptual loss to keep low-dose images from looking over-smoothed, with code. |
| CTformer: Convolution-Free Token2Token Dilated Vision Transformer for Low-Dose CT Denoising · code | PMB | 2023 | A convolution-free vision transformer for low-dose CT denoising, with code; a good representative of transformer-based methods. |
| LoDoPaB-CT, a benchmark dataset for low-dose computed tomography reconstruction · code | Scientific Data | 2021 | The descriptor of the LoDoPaB-CT benchmark dataset, with reference code; read it before reporting results on it. |
| Self-Supervised Training For Low-Dose CT Reconstruction | ISBI | 2021 | Trains a low-dose reconstruction network from the noisy measurements alone, without full-dose targets. |
Related pages
- Low-Dose CT Denoising
- Self-Supervised & Untrained Methods
- CT Reconstruction Datasets and Benchmarks
- Sparse-View CT Reconstruction: Methods, Datasets and Where to Start
- How to Choose a CT Reconstruction Dataset
Frequently asked questions
What is the standard benchmark for low-dose CT denoising?
The Mayo Clinic data from the 2016 AAPM Low-Dose CT Grand Challenge, which established the quarter-dose / 10% / 5% protocol; nearly every low-dose CT paper still reports on that split. The data are now released as LDCT-and-Projection-data on TCIA. LoDoPaB-CT is the main simulated benchmark.
Which low-dose CT datasets include projection data?
Among the low-dose datasets with a page here, these ship projection data: LoDoPaB-CT, LDCT-and-Projection-data, COBRA2026, Walnut-21 (two dose levels). The low-dose datasets table shows the rest.
Is LoDoPaB-CT enough on its own?
It is simulated from LIDC/IDRI chest scans at a single noise level (N0 = 4096), so other dose levels have to be re-simulated. Report on a real-data benchmark such as the Mayo data as well.
Can I train a low-dose CT denoiser without full-dose images?
Yes. Self-supervised methods avoid the need for paired full-dose images. Noise2Noise popularised training a denoiser without clean targets, and on CT it became the standard no-clean-ground-truth evaluation protocol.