Deep learning-based restoration of PET images from a dual-panel breast dedicated scanner.

Positron Emission Tomography (PET) is important for breast cancer diagnosis and monitoring, but high costs restrict access. Dual-panel scanners can reduce costs, though they typically produce lower quality images with quantitative bias compared to full-ring systems. In this study, we investigated the use of deep learning (DL) to address these limitations and improve image quality in a dedicated dual-panel breast PET scanner. Monte Carlo simulations were performed with the GATE toolkit to model both dual-panel and full-ring scanners. The dual-panel configuration included two detector heads separated by 21 cm, each consisting of 3×4 blocks of 13×13 crystals, while the full-ring system comprised 14 detector blocks in four rings with a 21 cm diameter. During acquisition, the dual panel system was rotated by 90 degrees (step and shoot, no data acquisition during motion) to increase angular sampling. Clinical data from 51 F-FDG breast PET/CT cases were used as activity and attenuation maps for the simulations. A SwinUNETR architecture was trained to synthesize full-ring-equivalent images from dual-panel data. Performance was evaluated with structural similarity index (SSIM), peak signal-to-noise ratio (PSNR), root mean square error (RMSE), and voxelwise correlation. The dual-panel and full-ring scanners achieved spatial resolutions of 3.2 mm and 1.6 mm, and sensitivities of 8.9 and 14.2 cps/kBq, respectively. Compared with dual-panel images, AI-enhanced outputs showed improvements of 2.65% in PSNR, 26.4% in SSIM, and 12.1% in RMSE. Voxelwise correlation increased markedly (R increased from 0.75 to 0.96). These findings highlight the potential of DL-based approaches to generate higher quality, artifact-reduced breast PET images, allowing cost-effective dual-panel systems to approach the performance of full-ring scanners.