Work

Sparse-View CT Reconstruction Using Deep Learning

Year: 2026
Type: preprint
Abstract: Sparse-view computed tomography (CT) reduces radiation dose and scan time by acquiring few projection angles but filtered back-projection (FBP) of such data suffers from severe streak artefacts because the angular sampling falls far below the Nyquist requirement. We study the post-processing approach in which a residual U-Net removes these artefacts from the FBP image, and test it on five synthetic datasets that differ in object class and physics: head phantoms at normal dose, chest phantoms at ten-fold lower dose, aluminium machined parts, photon-starved titanium components, and aluminium resolution targets. Sinograms at 30, 60 and 90 views are simulated with the Beer-Lambert law, Poisson photon statistics and Gaussian read noise, and the network is compared against FBP and the Simultaneous Algebraic Reconstruction Technique (SART), both with settings tuned on validation data. On 150 held-out test images per dataset, the U-Net achieves the highest peak signal-to-noise ratio (PSNR) in all 15 dataset and view settings, improving on the best classical method by 1.9 to 5.8 dB (paired Wilcoxon test, p < 10⁻²⁰ in every setting), and at 30 views it exceeds the PSNR of 180-view FBP on all five datasets, while running in about 20 ms per image on a CPU. Stress tests show the limits of the approach: networks transfer between similar industrial datasets but degrade or harm images of other object classes, fail when the test data are more sparsely sampled or receive ten times fewer photons than in training, and are outperformed by SART on the unseen Shepp-Logan phantom. A network trained on resolution targets does generalise to a resolution plate taken from laboratory data, improving on SART by 2.7 to 3.4 dB. The code, data and trained models are released for full reproducibility.
Language: English
Field: Medicine
Open Access status: green
OpenAlex ID: W7236349072
Cites: 12
Cited by:
Related to: 10