Spatiotemporal SPRINTER Framework Enhances Dynamic Brain PET Image Quality and Kinetics

A newly developed spatiotemporal denoising framework termed SPRINTER substantially enhances image quality and preserves tracer kinetic fidelity in four-dimensional dynamic brain positron emission tomography without relying on clean reference standards, according to an original research evaluation published in EJNMMI Physics by investigators at Washington University School of Medicine. Dynamic four-dimensional PET provides continuous in vivo monitoring of radiotracer distributions across targeted regions of interest over time, delivering essential quantitative insights into physiological mechanisms, disease progression, and therapeutic responses. However, dynamic scanning protocols divide administered activity into rapid sequential acquisitions, causing early and low-count time frames to suffer from severe statistical Poisson noise that limits downstream clinical utility and quantitative precision. By combining anatomically guided adaptive temporal filtering, supervised deep learning spatial feature extraction, and iterative Landweber-type regularization, the researchers demonstrated marked gains in signal-to-noise ratio, structural consistency, and kinetic parameter estimation across synthetic simulations and real human brain datasets involving five distinct radiotracers.

Dynamic PET data acquisition presents unique image processing hurdles that conventional static reconstruction pipelines cannot adequately resolve. Standard iterative reconstruction techniques, including ordered-subsets expectation maximization and maximum likelihood expectation maximization, are designed to address emission noise characteristics but frequently produce substantial statistical fluctuations when applied to brief dynamic intervals. Although deep neural networks have emerged as a promising tool for biomedical noise reduction, most architectures rely on supervised training regimes that require large libraries of paired noisy and pristine ground-truth images. While synthetic phantom simulations can supply artificial ground truth, artificial training data frequently introduce domain shift discrepancies that degrade neural network performance when transferred to complex clinical brain scans acquired under varying scanner configurations and injection protocols.

Alternative unsupervised filtering and machine learning methods designed to circumvent paired training requirements introduce their own significant algorithmic compromises. Classical post-processing algorithms such as non-local means filtering and bilateral filtering rely primarily on spatial patch averaging and intensity similarities, suppressing noise across static structures but neglecting temporal inter-frame dependencies. Consequently, these spatial filters cause pronounced spatial oversmoothing and distort rapid temporal tracer kinetics, leading to inaccurate time activity curves, particularly in early dynamic frames where count densities are lowest. Meanwhile, modern unsupervised deep learning models such as Noise2Void operate on single noisy frames by masking random voxels and predicting their values from spatial neighbors. Although Noise2Void bypasses the need for paired clean targets, its purely spatial dependency structure lacks temporal awareness across adjacent time frames. Advanced spatiotemporal filtering algorithms such as Kinetics-Induced Block Matching 5D and Spatiotemporal Expectation-Maximization also encounter severe hurdles, including extensive manual parameter tuning or aggressive temporal averaging that induces systematic kinetic bias.

To overcome these collective processing limitations, the investigators engineered SPRINTER around an anatomically guided, self-supervised adaptive principal component analysis routine designed to perform temporal smoothing without distorting underlying physiological trajectories. Dynamic PET list-mode acquisitions were coupled with high-resolution structural T1-weighted magnetization-prepared rapid gradient-echo magnetic resonance imaging scans, which provided automated tissue segmentation through FreeSurfer tools into eight designated regions of interest. These anatomical compartments comprised cerebral white matter, cerebral gray matter, cerebellar white matter, cerebellar gray matter, cerebrospinal fluid, vascular structures, the brainstem, and non-brain tissue. Within each segmented volume, voxel time activity curves were isolated into matrix representations and decomposed using principal component analysis, under the biological premise that voxels within a single anatomical compartment share coherent radiotracer uptake behaviors. To accommodate fluctuating noise across framing intervals, the optimal number of principal components was dynamically adjusted via a cost function that constrained regional curve reconstruction error strictly below five percent before applying a temporal moving average filter.

Because patient-specific anatomical adaptive principal component analysis requires considerable computational time and can introduce edge artifacts along segmented boundaries, the investigators utilized the smoothed outputs as self-supervised training targets for a lightweight three-dimensional patch-based convolutional neural network. The team designed a 3D ResUNet architecture incorporating dual input channels consisting of coregistered T1-weighted structural MRI volumes alongside time-sampled three-dimensional dynamic PET frames. The network architecture featured seven convolutional layers across its contracting path using kernel sizes of seven, five, three, and one paired with leaky rectified linear unit activations, and seven corresponding expansion layers utilizing kernel sizes of one, three, seven, and nine paired with standard rectified linear unit activations. Optimization was executed using the Adam algorithm with a batch size of ten, an initial learning rate of 0.001 halved every 50,000 batches, and an objective function minimizing the L1 norm between network outputs and the adaptive principal component targets, reducing reconstruction latency from roughly 200 seconds down to several seconds per dynamic frame.

Although the trained 3D ResUNet rapidly suppressed high-frequency spatial noise across dynamic frames, the raw neural network outputs exhibited structural oversmoothing and degraded temporal signal-to-noise ratios, necessitating a dedicated temporal restoration phase. To restore dynamic fidelity without reintroducing spatial blurring kernels, SPRINTER performs temporal principal component decomposition on the spatially denoised outputs, isolating the leading five principal components that account for at least ninety-five percent of overall temporal variance. The original noisy ordered-subsets expectation maximization dynamic sequence was then projected across these denoised temporal components while referencing the noisy regional mean signal to prevent kinetic offset bias. By restricting temporal synthesis to the leading five components, the temporal operator selectively preserved dominant tracer kinetic variations while suppressing higher-order components dominated by Poisson counting noise. The resulting voxel-wise weight maps mapped tissue-specific temporal variations across cerebral anatomy, providing an interpretable representation of dynamic radiotracer uptake without relying on arbitrary spatial blurring filters.

The individual spatial and temporal outputs were ultimately integrated into an iterative spatiotemporal regularization pipeline that refined dynamic brain volumes through a Landweber-type deconvolution scheme. At each iteration, the algorithm computes a residual term enforcing consistency with the fixed spatially denoised ResUNet volume alongside a correction term derived from the temporal denoising operator. Rather than relying on fixed hyperparameter tuning, SPRINTER automatically calculates data-driven adaptive step sizes at every step by normalizing residual and correction projections against frame-specific noise levels measured as normalized image standard deviations. During initial iterations, residual step sizes remain elevated to preserve spatial stability, while correction step sizes gradually increase in subsequent iterations to enforce temporal kinetic accuracy. The iterative process terminates when relative updates between successive iterations fall below 0.0001 or reach a ceiling of twenty iterations, with practical convergence typically occurring within ten to twelve iterations and completing full forty-frame patient reconstructions within 300 to 400 seconds.

To rigorously quantify denoising efficacy against verifiable ground-truth standards, the authors generated realistic synthetic four-dimensional PET simulations using an analytical simulation environment derived from high-resolution structural brain atlases containing sixty-one distinct anatomical regions. Simulated regional time activity curves were assigned realistic activity concentrations in becquerels per liter, forward-projected into sinogram matrices, attenuated using MRI-derived attenuation maps, and subjected to Poisson counting statistics calibrated to typical injected clinical activities between 700 and 950 megabecquerels. Detected true coincidence event totals accurately replicated real dynamic acquisitions, scaling from approximately 100,000 counts in early transient frames to approximately 10,000,000 counts in late static intervals. Downstream kinetic analysis was executed on simulated Fluorine-18 fluorodeoxyglucose sequences using the EMATA toolbox to compute influx constant Ki parametric maps across cerebral white and gray matter structures, enabling direct evaluation of how denoising pipelines influence physiological parameter estimation relative to known ground-truth maps.

Simulation evaluations demonstrated that SPRINTER outperformed baseline ordered-subsets expectation maximization, spatial-only, temporal-only, non-local means, and Noise2Void methods across all tested quantitative metrics. Relative to classical non-local means filtering, SPRINTER achieved significant reductions in normalized root mean square error, dropping from 0.58 to 0.09 for Fluorine-18 fluorodeoxyglucose, from 0.42 to 0.07 for Fluorine-18 florbetapir, from 0.27 to 0.12 for Carbon-11 Pittsburgh compound B, from 0.33 to 0.07 for Oxygen-15 water, and from 0.37 to 0.09 for Oxygen-15 oxygen. Structural similarity index measurements showed corresponding improvements over non-local means, climbing from 0.17 to 0.73 for fluorodeoxyglucose, 0.19 to 0.83 for florbetapir, 0.27 to 0.57 for Pittsburgh compound B, 0.25 to 0.73 for Oxygen-15 water, and 0.25 to 0.64 for Oxygen-15 oxygen. Downstream kinetic modeling on simulated fluorodeoxyglucose revealed that SPRINTER reduced whole-brain Ki parametric normalized mean square error from 1.10 down to 0.43 compared to non-local means, representing an approximate sixty percent error reduction and avoiding the structural biases observed with spatial-only smoothing.

Experimental validation on real human brain scans encompassed 220 study participants enrolled under institutional review board approval who underwent dynamic neuroimaging on a simultaneous Biograph mMR PET/MRI scanner across five radiotracers, with Carbon-11 Pittsburgh compound B withheld during network training to evaluate cross-tracer generalizability. In acquired patient datasets, SPRINTER achieved the highest signal-to-noise ratio across all radiotracers, elevating values from 2.12 to 2.76 in Fluorine-18 fluorodeoxyglucose, 3.95 to 4.33 in Fluorine-18 florbetapir, 3.05 to 3.35 in Carbon-11 Pittsburgh compound B, 2.68 to 3.92 in Oxygen-15 water, and 2.53 to 3.71 in Oxygen-15 oxygen compared to non-local means. Contrast-to-noise ratios between cerebral white and gray matter increased significantly for all tracers except Oxygen-15 oxygen, while perception-based image quality evaluator scores improved by an average of 20.5 percent relative to baseline reconstruction. Furthermore, analysis of normalized time activity curve errors across sixty-one anatomical brain regions confirmed that SPRINTER maintained regional kinetic consistency without generating the edge-blurring distortions seen in spatial baselines.

Despite these documented performance advantages, the investigators highlighted several clear methodological constraints and operational limitations that warrant consideration prior to broader clinical translation. First, the self-supervised adaptive principal component analysis module depends directly on automated structural FreeSurfer segmentation of co-registered T1-weighted MRI volumes, rendering the workflow vulnerable to inter-modality spatial misalignments and computationally demanding segmentation runtimes that could introduce boundary artifacts. Second, although iterative regularization markedly mitigates oversmoothing, the spatial 3D ResUNet component can still induce subtle residual blurring within early low-count intervals characterized by profound Poisson noise. Third, because human validation was conducted retrospectively using imaging data acquired on a single PET/MRI scanner platform, multi-center studies will be required to confirm generalizability across diverse scanner hardware, acquisition geometries, and reconstruction protocols. Finally, quantitative tracer kinetic parameter validation was established exclusively within simulated dynamic datasets with known mathematical ground truth, emphasizing the necessity of conducting future clinical validation incorporating arterial blood sampling or established reference standards.

Subject of Research: Self-supervised spatiotemporal denoising for four-dimensional dynamic brain PET imaging

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

Article References: Yousefi, H., Hamdi, M., Wang, Y., Wang, B., Laforest, R., Brier, M., Chen, Y., & An, H. (2026). Denoising of 4D dynamic PET images using spatiotemporal regularization with integrated temporal restoration (SPRINTER). EJNMMI Physics, 13(1), Article 94. https://doi.org/10.1186/s40658-026-00898-w

Image Credits: AI Generated

DOI: 10.1186/s40658-026-00898-w

Keywords: dynamic PET, spatiotemporal denoising, SPRINTER, positron emission tomography, PET-MRI, adaptive PCA, 3D ResUNet, deep learning, kinetic modeling, Ki parametric imaging, neuroimaging, image reconstruction, noise reduction, time activity curves, 18F-FDG, 18F-AV45, 11C-PiB, 15O-water