Self-Supervised & Untrained Methods
Self-supervised and untrained methods reconstruct or denoise CT images without paired clean ground truth. They exploit structure in the measurements themselves, such as independent noise in repeated or split data, consistency with the measured sinogram or invariance to transformations, or the implicit bias of an untrained network.
They matter because clean, paired training data are scarce in CT: a patient cannot be scanned twice at full and low dose for research, metal-free versions of implant scans do not exist, and industrial and synchrotron data rarely come with references.
The main families are Noise2Noise-style denoising trained on noisy data alone; the deep image prior and other untrained networks fitted to a single scan; per-scan neural fields that represent the volume and fit the measured projections; equivariant-imaging losses; and hybrid supervised-unsupervised schemes. They avoid training-set bias but are usually slower at test time and need careful early stopping or regularisation.
Browse all 35 Self-Supervised & Untrained Methods papers in the searchable directory
Start here
If you are new to this topic, these are good first reads, chosen by the maintainers.
- Computed Tomography Reconstruction Using Deep Image Prior and Learned Reconstruction Methods (Inverse Problems, 2020 · code)
Applies the deep image prior to low-dose and sparse-angle CT with small training sets, with a benchmark codebase. - 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. - IntraTomo: Self-supervised Learning-based Tomography via Sinogram Synthesis and Prediction (ICCV, 2021 · code)
Represents the density field as a network fitted to an incomplete sinogram, refined with geometric priors, with code. - Unified Supervised-Unsupervised (SUPER) Learning for X-ray CT Image Reconstruction (TMI, 2021)
Unifies supervised and unsupervised learning inside model-based iterative reconstruction.
Papers
The 30 most recent of 35 papers, newest first.
Browse all 35 Self-Supervised & Untrained Methods papers in the searchable directory
Datasets, benchmarks and toolkits
Entries from the directory whose tasks or tags match this topic.
- Walnut-21 (two dose levels) (official page) · Dataset — Projection only. High-resolution cone-beam scans of 21 walnuts, each acquired at two radiation dose levels - a good and a noisy version of the same object.
- Noise2Noise denoising protocol (official page) · Benchmark — The reference implementation that popularised training a denoiser without clean targets.
- tomosipo · Toolkit — Painless specification of 2D/3D projection geometries; wraps ASTRA and integrates with PyTorch.
- torch-radon · Toolkit — Differentiable Radon transform and filtered backprojection in PyTorch.
Guides
Guides that cover this topic.
Other topics
- Surveys & Reviews (24)
- Classical & Iterative Reconstruction (71)
- Sparse-View CT (157)
- Limited-Angle Tomography (46)
- Low-Dose CT Denoising (69)
- Metal Artifact Reduction (23)
- Spectral & Dual-Energy CT (19)
- Cone-Beam CT (77)
- Dynamic & 4D CT (24)
- Scatter Correction (11)
- Geometry Calibration & Motion Compensation (28)
- Deep Unrolling & Model-Based Learning (68)
- Diffusion & Generative Priors (68)
- Implicit Neural Representations (25)
- Gaussian Splatting (26)