Deep learning-guided late-phase dynamic frame prediction for accelerated Patlak analysis in ¹⁸F-FDG-PET brain imaging

Dynamic ¹⁸F-FDG-PET enables the quantitative assessment of cerebral glucose metabolism but requires prolonged acquisition times, which pose challenges to patients presenting with neurodegenerative disorders. This study aims to use deep learning to reduce dynamic FDG-PET scan duration while preserving quantitative kinetic modeling accuracy by predicting late-phase PET frames from intermediate-phase data.

A deep learning framework based on a 3D U-Net was developed to generate the final six late-phase dynamic PET frames (frames 47-52) using intermediate-phase frames (frames 34-46) as input. The proposed DL model was benchmarked against two training-free temporal approaches: linear least-squares (LLS) extrapolation and a Hermite/Pitch-In-based (Hermite) method. The model was trained and evaluated on dynamic brain FDG-PET datasets. Image-level similarity between DL-predicted and reference images was assessed using the structural similarity index (SSIM) and peak signal-to-noise ratio (PSNR). Quantitative reliability was evaluated using Patlak reference tissue modeling with the cerebellum as the reference region and a linear phase start time of 25 min post-injection. The influx rate constant (K_i) and intercept (V₀) were estimated across six predefined brain regions for both DL-predicted and original images, serving as standard of reference.

The DL-predicted images showed excellent whole-brain agreement with reference images, achieving a mean SSIM of 0.93 and PSNR of 33.8 dB. Quantitative error was low, with a whole-brain root mean square error (RMSE) of approximately 0.16 standardized uptake value ratio (SUVR) and negligible bias (mean error (ME) ≈ 0). Compared with the LLS and Hermite training-free approaches, the proposed DL model achieved higher image-level fidelity and more accurate regional Patlak quantification, with K_i and V₀​ estimates showing closer agreement with the reference values across all six brain regions.

The proposed DL model enables a reduction of dynamic FDG-PET acquisition time from 60 to 30 min while preserving accurate quantification of kinetic parameters. This method improves patient comfort and scanner throughput without compromising the reliability of Patlak-derived kinetic parameters, thereby supporting its potential for clinical and research applications in dynamic brain PET imaging.