Denoising of 4D dynamic PET images using spatiotemporal regularization with integrated temporal restoration (SPRINTER).

Abstract

Purpose: High-quality 4D dynamic PET imaging is often compromised by noise, especially in low-count frames, which limits clinical utility and quantitative accuracy. This study proposes a novel spatiotemporal denoising method (SPRINTER) that integrates anatomical priors, self-supervised adaptive principal component analysis (aPCA), and deep learning to enhance image quality and preserve time activity fidelity without relying on ground-truth data.

Methods: The method combines anatomically guided aPCA-based temporal smoothing with a supervised 3D ResUNet for spatial denoising. The resulting high signal-to-noise ratio, spatially denoised images are iteratively refined using a temporal regularization strategy to enhance contrast. The model was trained using brain scan datasets from 220 participants who were administered one of the following PET/MRI radiotracers: ¹⁸F-AV45, ¹⁸F-FDG, ¹⁵O-HO, ¹⁵O-OO, and ¹¹C-PiB for generalizability testing. Evaluations were performed on synthetic simulations with known ground-truth data and real acquired brain scan data. We compared the proposed method against multiple baseline approaches, including conventional reconstruction (OSEM), spatial-only and temporal-only denoising strategies, and two unsupervised reference methods: Non-Local Means (NLM) filtering, a classical denoising approach, and Noise2Void (N2V). These baselines were selected to reflect both conventional and recent unsupervised deep learning denoising paradigms applicable to PET imaging. Performance was assessed using both synthetic simulations with known ground truth and real acquired PET data. In addition to voxel-level image quality metrics, the impact of denoising on downstream quantitative analysis was evaluated using Ki parametric imaging derived from simulation data. This enabled direct assessment of tracer kinetic estimation accuracy across different denoising strategies.

Results: We compared SPRINTER with OSEM, spatial-only, temporal-only, and two existing unsupervised denoising approaches, including NLM and Noise2Void. Across both simulation and acquired datasets, SPRINTER significantly outperformed all comparison methods (p < 0.05, ANOVA with Bonferroni multiple comparison corrections). The largest gains were observed for the noisiest data using ¹⁵O-HO and ¹⁵O-OO tracers, where SPRINTER achieved marked SNR improvements in acquired data (e.g., ¹⁵O-HO: 3.92 ± 1.28 vs. 2.68 ± 1.24; p = 0.096; ¹⁵O-OO: 3.71 ± 1.79 vs. 2.53 ± 1.15 for NLM; p = 0.082). These improvements were consistently superior to those obtained with NLM and Noise2Void, which showed either spatial blurring or limited performance in low-count frames. In simulation studies, SPRINTER significantly reduced voxel-wise reconstruction error compared to all baselines (NRMSE: 0.07 ± 0.04 and 0.09 ± 0.05 vs. 0.33 ± 0.08 to 0.07 and 0.37 ± 0.31 for NLM; p < 0.0001), while also outperforming NLM and Noise2Void, which exhibited higher residual error and reduced temporal consistency. Importantly, contrast was preserved and enhanced, with improved CNR particularly in ¹⁸F-FDG and ¹¹C-PiB (e.g., ¹⁸F-FDG: 0.37 ± 0.33 vs. 0.18 ± 0.19; p < 0.0001; ¹¹C-PiB: 0.54 ± 0.08 vs. 0.41 ± 0.12; p < 0.001), indicating effective denoising without loss of biologically relevant signal. While NLM and spatial-only methods achieved comparable SNR improvements, this was largely driven by oversmoothing. In contrast, with a similar level of SNR enhancement, the incorporation of temporal information further improved structural fidelity, as reflected by higher SSIM values across all tracers in simulation data (e.g., ¹⁸F-AV45: 0.73 ± 0.14 vs. 0.17 ± 0.08 for NLM; p < 0.0001).

Conclusion: This study introduces a robust, self-supervised spatiotemporal denoising method for 4D dynamic PET imaging of the brain. SPRINTER spatiotemporal denoising leverages anatomical guidance and temporal dynamics to enhance spatial clarity and temporal consistency. Its radiotracer-agnostic design and rigorous self-supervision make it well-suited for clinical applications where high-quality reference data are unavailable.