Sparse-View CT Reconstruction
Physics-guided deep learning for reconstructing CT volumes from a small number of projections — my undergraduate final-year project, carried out as a research intern at A*STAR I2R.
- Category
- Machine Learning
- Published
- May 2026
- Stack
- PyTorch
- 3D U-Net
- VGGT
- Transformer
- Singularity
- NSCC HPC
The problem
Computed tomography normally needs projections from hundreds of angles. Acquiring fewer views lowers radiation dose and scan time, but classical filtered back-projection (FBP/FDK) then produces heavy streak artefacts. Reconstruction becomes an ill-posed inverse problem.
What I did
- Built the full pipeline in PyTorch: data preprocessing, model training and quantitative evaluation.
- Implemented and ablated three architectures — 3D U-Net, VGGT and a Transformer — integrating physics priors (filtered back-projection constraints) with learned components.
- Ran three ablation studies on the NSCC HPC cluster in Singularity containers, tuning learning-rate schedules and loss-function combinations.
Result
With 30 projections, the best model improved on the conventional FDK baseline by +10 dB PSNR and +0.5376 SSIM. The work was delivered with a mid-term report and a video defence.
Context
A*STAR Institute for Infocomm Research (I2R), Singapore · Jan – May 2026.