Time-aware frame-aligned diffusion-inspired model for late dynamic PET frame prediction from early [¹¹C] raclopride

To reduce the burden of dynamic PET acquisition while preserving quantitative analysis, we propose a time-aware frame-aligned diffusion-inspired residual denoising framework for predicting late dynamic PET frames from a short early acquisition window (10 min in this study) of [¹¹C] raclopride.

Irregular native dynamic frames were converted to a uniform 1-min indexing grid using overlap-weighted rebinning. The first 10 min were used as the early input and to derive the spatial and intensity preprocessing parameters. For each requested late time, a shared time-aware 3D U-Net predicted an anchor-relative residual with respect to the last available early frame. A training-only DDPM-inspired Gaussian perturbation and noise-prediction branch shared the prediction backbone and regularized residual learning, while temporal-consistency loss constrained predicted frame-to-frame changes. All learning-based methods used identical subject-wise five-fold cross-validation within the development cohort, followed by evaluation on a fixed internal test set; the external cohort was held out from model selection.

The proposed approach achieved higher PSNR and SSIM and lower MAE than the comparison methods, while providing closer agreement with reference TACs and SRTM-derived parameter estimates across the limbic, associative, and sensorimotor striata.

The proposed framework provides deterministic time-resolved late-frame estimates from early dynamic PET while using a DDPM-inspired noise-prediction objective only as an auxiliary training regularizer rather than as an iterative diffusion inference process. This design may support shortened dynamic PET protocols while retaining TAC and kinetic-parameter fidelity.