Deep learning-based one-stop ¹¹C-CFT and ¹⁸F-FDG dual-tracer brain PET imaging protocol for Parkinson's disease

This study explored using deep learning (DL) to separate short-interval staggered ¹¹C-CFT/¹⁸F-FDG brain PET images for Parkinson's disease, aiming to reduce scan waiting time. A total of 67 patients performing ¹¹C-CFT and ¹⁸F-FDG brain PET scans on separate days were retrospectively enrolled. A Swin UNETR model was trained to generate pseudo ¹⁸F-FDG PET images from simulated dual-tracer sum images. The simulation assumed ¹⁸F-FDG was administered at 80, 100, 120, or 200 min (∆t) after ¹¹C-CFT injection. Compared to actual ¹⁸F-FDG images, the pseudo images showed high visual similarity across all ∆t intervals. Low average NMSE values (∼0.0004) and high average SSIM values (0.9991-0.9993) were consistently achieved across all Δt groups. Bland & Altman analysis of the whole brain region demonstrated low average SUVR bias across all Δt groups remained within ±0.001. Region-wise correlation analysis revealed strong correlations between actual and pseudo ¹⁸F-FDG images across all Δt, with slopes ranging from 0.994 to 1.001, and all R²> 0.99. SUVmean, LBR and SNR values for pseudo ¹⁸F-FDG images exhibited no statistically significant differences compared to actual ¹⁸F-FDG images (P > 0.05). Dual-tracer PET images can be effectively separated using the DL model, yielding high-quality visual and semi-quantitative results when ¹⁸F-FDG is injected immediately after the ¹¹C-CFT PET scan, thereby reducing patient wait time, improving patient comfort, and enhancing overall clinical efficiency.