Image Reconstruction Techniques in Computed Tomography

Computed tomography (CT) reconstructs volumetric images by mathematically inverting two-dimensional X-ray projections acquired around an object. Historically, reconstruction has been dominated by analytical methods such as filtered back projection (FBP), valued for its computational efficiency but sensitive to noise and limited-angle data. Iterative reconstruction techniques, meanwhile, solve a system of equations through repeated updates, offering improved image fidelity by modelling noise statistics, system geometry and prior information. Recent advances have introduced statistical regularisation, compressed sensing and model-based algorithms, which mitigate radiation dose while preserving diagnostic detail. More recently, machine learning and deep neural networks have been integrated into reconstruction pipelines to enable sparse-view imaging, artefact suppression and accelerated processing. Hybrid approaches, combining physics-based models with data-driven priors, are now enabling high-resolution, low-dose and real-time applications across medical, industrial and synchrotron imaging. These technical innovations hold promise for global health, non-destructive testing and dynamic studies in materials science.

Research from Nature Portfolio

Recent studies have demonstrated deep residual learning to suppress streak artefacts in sparse-view CT. A U-shaped convolutional network was trained to learn the mapping between artefact-corrupted reconstructions and full-view images, effectively restoring structural detail and contrast. This approach achieves image quality close to that of conventional full-dose protocols while significantly reducing projection requirements, paving the way for faster scans and lower radiation exposure.

Research from all publishers

A deep neural framework has been applied to nanoscale X-ray tomography, combining ptychographic modelling with a three-dimensional U-net to reconstruct high-resolution images from limited projections, enabling rapid inspection of integrated circuits. In synchrotron facilities, a full-stack pipeline powered by deep learning has been proposed to handle petabyte-scale datasets, encompassing denoising, artefact correction and tomographic inversion in real time. In clinical low-dose CT, a convolution-free dilated vision transformer has been introduced for denoising, utilising long-range attention to enhance edge definition and structural coherence. This model outperforms state-of-the-art CNNs at minimal computational cost, making it suitable for routine diagnostic workflows.

Image Reconstruction Techniques in Computed Tomography publication trend

The graph below shows the total number of articles in image reconstruction techniques in computed tomography across all publications each year (not limited to Nature Index journals).

Technical terms

Filtered back projection (FBP): An analytical method that reconstructs images by filtering and back-projecting projection data.

Iterative reconstruction: A numerical technique that refines image estimates through repeated forward and back-projection steps, often incorporating noise models and regularisation.

Deep learning: A set of machine learning methods using multi-layer neural networks to learn complex mappings from data.

U-net: A convolutional neural network architecture with symmetric encoding and decoding paths, widely used for image restoration and segmentation.

Sparse-view CT: Imaging acquired with a reduced number of projections to minimise dose or acquisition time, often requiring advanced reconstruction to suppress artefacts.

Vision transformer: A neural network architecture using self-attention mechanisms to capture long-range dependencies in image data.

References

1. Deep learning enables nanoscale X-ray 3D imaging with limited data. Light: Science & Applications (2023).
2. Towards full-stack deep learning-empowered data processing pipeline for synchrotron tomography experiments. The Innovation (2023).
3. Artifact Removal using Improved GoogLeNet for Sparse-view CT Reconstruction. Scientific Reports (2018).
4. CTformer: convolution-free Token2Token dilated vision transformer for low-dose CT denoising. Physics in Medicine and Biology (2023).