Projection restoration with dual-domain Swin refinement for sparse-view CT reconstruction

We investigated whether projection restoration before reconstruction, followed by image-domain refinement, improves sparse-view computed tomography (CT) quality under aggressive undersampling.

A compact dual-domain Swin framework restores sparse sinograms before geometry-aware differentiable filtered back projection (FBP) and refines the resulting images. The projection branch uses gated Swin processing and angular upsampling with a bicubic bypass. Abdominal AAPM sinograms contained 2304 full views, retaining 576, 288, and 144 views at 4x, 8x, and 16x. Comparators were FBP, bilinear and bicubic interpolation, ProCT, and geometry-adapted DuDoTrans; ablations used ten epochs.

DuDoSwin achieved PSNR/SSIM/LPIPS of 45.64 dB/0.981/0.011 at 4x, 44.11 dB/0.975/0.020 at 8x, and 42.02 dB/0.963/0.035 at 16x. Mean PSNR and SSIM exceeded ProCT and DuDoTrans; DuDoTrans had lower LPIPS at 4x and 8x. DuDoSwin used 1.36-1.80 million parameters but was slower than both learned comparators. Comparisons against interpolation and DuDoTrans were significant in Holm-corrected paired tests. Bypass removal improved ablation PSNR by 0.35 dB at 8x. Evaluation included the validation patient; exclusion changed mean PSNR by at most 0.1 dB and altered no ranking. Slice-wise tests ignore within-patient dependence.

Explicit projection restoration followed by image-domain refinement yielded favorable PSNR and SSIM, with perceptual metric and computational trade-offs. Neither residual shortcut improved the 8x ablation results. Independent patient evaluation remains necessary.