Enhancing uncertainty assessment in dynamic PET imaging with residual permutation and clustering

Quantitative positron emission tomography (PET) is widely used for disease diagnosis and therapy monitoring, yet the reliability of kinetic parameters depends on robust uncertainty quantification. Existing Bayesian methods are computationally demanding, bootstrap approaches are noise-sensitive, and recent deep learning models often require large training datasets and lack physical scale sensitivity. To address these limitations, we propose a clustering-based residual permutation (RP) framework for uncertainty estimation in dynamic PET. The method generates pseudo time-activity curves (TACs) by permuting fitting residuals within kinetically homogeneous clusters, preserving spatiotemporal noise characteristics while avoiding noise misallocation across heterogeneous regions. To ensure meaningful residual construction, we introduce a regularized model with Huber loss and elastic-net regularization, improving numerical stability and preventing overfitting. Extensive validation on simulated data (TACs and XCAT-OSEM reconstructions) and clinical total-body PET demonstrates that RP yields uncertainty estimates that scale consistently with noise level and preserve expected physical disparities between kinetic parameters. Compared with reference baselines, the proposed framework provides a distribution-free, training-independent, and computationally efficient solution for voxel-wise uncertainty quantification. Overall, RP fills a practical gap between expensive Bayesian inference and data-hungry deep learning, offering a robust and clinically deployable approach to uncertainty-aware dynamic PET analysis. The code of the proposed RP method and reference methods is available at: https://github.com/ANMMILab-PKU.