Dual-domain transformers enhance PET primal-dual reconstruction

Transformer-based LPD reconstruction outperforms CNN baseline in PET: Stacked Transformer and Restormer improve MSE and PSNR across noise levels; Restormer yields higher SSIM at low noise. Models trained only on synthetic data generalized to experimental measurements.

• Cross-attention UNet LPD improved MSE and PSNR over baseline but slightly reduced SSIM, indicating a trade-off between pixel-wise accuracy and structural similarity.
• A system-aligned synthetic data generation process mimicking a preclinical PET scanner enabled robust transfer learning to real measurements.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.