Dual Domain Swin Transformer based Reconstruction method for Sparse-View Computed Tomography
A new sparse-view parallel beam computed tomography reconstruction method is proposed that exploits the restoration capabilities of Transformer networks, in particular the Swin Transformer-based image reconstruction network SwinIR. Our method comprises three key blocks: sinogram upsampling via line...
Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Article |
| Language: | deu |
| Published: |
NDT.net
2025-02-01
|
| Series: | e-Journal of Nondestructive Testing |
| Online Access: | https://www.ndt.net/search/docs.php3?id=30751 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | A new sparse-view parallel beam computed tomography reconstruction method is proposed that exploits the restoration capabilities of Transformer networks, in particular the Swin Transformer-based image reconstruction network SwinIR. Our method comprises three key blocks: sinogram upsampling via linear interpolation, initial reconstruction using deep learning in both domains, and residual refinement. Two architectures are tested: a long one using neural networks in both domains of the residual refinement block and a short one using a network exclusively in the sinogram domain. Each method is tested with SwinIR and UNet, resulting in four variants, all of which outperform traditional methods like FBP and SIRT in terms of PSNR and SSIM. The short architecture using SwinIR achieves the best results, with a training and computation time smaller than the SwinIR-based long architecture but larger than both U-Net-based variants.
|
|---|---|
| ISSN: | 1435-4934 |