Implicit Neural Representations
Implicit neural representations (INRs) represent the unknown CT volume as a coordinate-based neural network that maps a 3D (or 3D plus time) position to an attenuation value, and fit that network so that its simulated projections match the measured ones. The network is the image; reconstruction is per-scan optimisation rather than inference with a pre-trained model.
Because the representation is continuous and fitted to each scan, INRs need no training data and handle sparse-view, limited-angle and dynamic acquisitions naturally; multiresolution hash encodings make them fast enough for cone-beam volumes. Methods in this family are often called neural attenuation fields or NeRF-style X-ray methods.
Current work focuses on encodings and architectures that reduce over-smoothing or noise fitting, priors from pre-trained or diffusion models, motion fields for 4D-CT, polychromatic forward models for metal artifacts, and uncertainty quantification. Gaussian splatting is a closely related explicit alternative and has its own topic page.
Browse all 25 Implicit Neural Representations papers in the searchable directory
Start here
If you are new to this topic, these are good first reads, chosen by the maintainers.
- NAF: Neural Attenuation Fields for Sparse-View CBCT Reconstruction (MICCAI, 2022 · code)
Neural attenuation fields with hash encoding, the standard INR baseline for sparse-view CBCT, with code. - Structure-Aware Sparse-View X-Ray 3D Reconstruction (CVPR, 2024 · code)
A structure-aware neural radiodensity field with a line-segment transformer backbone, with code. - Dynamic CT Reconstruction From Limited Views With Implicit Neural Representations (ICCV, 2021)
Extends INRs to rapidly deforming scenes with a parametric motion field for 4D-CT. - Unsupervised Polychromatic Neural Representation for CT Metal Artifact Reduction (NeurIPS, 2023 · code)
A polychromatic INR that models metal physics for unsupervised metal artifact reduction, with code. - PINER: Prior-Informed Implicit Neural Representation Learning for Test-Time Adaptation in Sparse-View CT Reconstruction (WACV, 2023 · code)
Uses an INR for test-time adaptation of sparse-view CT models under unknown noise, with code.
Papers
All 25 papers, newest first.
Browse all 25 Implicit Neural Representations papers in the searchable directory
Datasets, benchmarks and toolkits
Entries from the directory whose tasks or tags match this topic.
- Finnish Inverse Problems Society (FIPS) datasets (official page) · Dataset — Projection only. A family of real measured sinograms (walnut, lotus root, apple, carrot, phantoms) with carefully documented acquisition geometry — the cleanest public source for testing reconstruction operators.
- FleX-ray Walnut CBCT Collection (CWI) (official page) · Dataset — Projection + image. 42 walnuts scanned on three different orbits each, so that artefact-free reference volumes can be computed from the combined views and used as supervised-learning ground truth.
- TIGRE · Toolkit — GPU-accelerated (CUDA) toolbox for cone-beam and parallel-beam CT; FDK, SART, SIRT, CGLS and TV variants.
- 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)
- Self-Supervised & Untrained Methods (35)
- Diffusion & Generative Priors (68)
- Gaussian Splatting (26)