Efficient 3D Whole-Body PET Image Denoising via Conditional Rectified Flow With Optimized Sampling Strategy

Positron Emission Tomography (PET) is essential for molecular imaging, but ultra-low-dose protocols—necessary to minimize radiation exposure—introduce significant noise that degrades diagnostic quality. While 3D diffusion models offer high reconstruction fidelity, their slow inference times hinder clinical adoption. This paper addresses the trade-off between image quality and computational efficiency by introducing a 3D rectified flow framework designed for rapid, high-quality whole-body PET denoising.

The authors propose a one-pass conditional 3D rectified flow (3D Flow) model. Unlike standard diffusion models that solve stochastic differential equations over many steps, this framework learns a deterministic vector field that transports Gaussian noise directly to a clean PET image along a linear trajectory. To optimize this process, the authors introduce a non-uniform sampling strategy that adjusts the integration schedule to prioritize regions of the trajectory where the model's velocity field is most complex. The model was trained on the UDPET dataset and evaluated using zero-shot transfer on an independent clinical dataset (FAHZU) across various dose levels.

The proposed 3D Flow model achieves competitive image quality (PSNR and SSIM) compared to 3D DDPM and DDIM baselines while reducing inference time from hours to approximately 30 seconds per volume. The non-uniform sampling strategy, specifically using two integration steps, significantly improves reconstruction fidelity compared to uniform sampling. Furthermore, the model demonstrates robust zero-shot generalization across unseen dose levels (down to 1/100 of the standard dose) and different clinical acquisition protocols, maintaining structural integrity and lesion conspicuity without requiring fine-tuning.

By enabling high-quality PET denoising in seconds, this framework bridges the gap between computationally intensive generative models and the throughput requirements of clinical workflows. This approach could facilitate the wider adoption of ultra-low-dose PET imaging, potentially reducing patient radiation exposure without compromising diagnostic confidence.