Evaluation of motion correction on myocardial [^{1 8}F]F‐AraG PET uptake using data‐driven respiratory gating

Abstract

Background

[^{1 8}F]F‐AraG (radio‐fluorinated 9‐β‐D‐arabinofuranosylguanine) is a PET radiotracer for imaging T‐cell activation with emerging applications in myocardial imaging. Cardiac PET is particularly susceptible to respiratory motion, which degrades spatial resolution and introduces regional variability in quantitative measurements.

Purpose

To evaluate the impact of data‐driven respiratory motion correction on myocardial image quality and quantitative stability in [^{1 8}F]F‐AraG PET within a practical whole‐body imaging workflow.

Methods

Respiratory motion correction was performed using a vendor‐implemented data‐driven gating approach (OncoFreeze AI, Siemens Healthineers), which derives motion signals from PET coincidence data without external hardware and applies phase‐based binning with motion‐compensated reconstruction. [^{1 8}F]F‐AraG PET scans from 15 healthy subjects were analyzed using two myocardial segmentation strategies: a threshold‐based method and a fixed‐size VOI approach. Quantitative metrics (SUVmean, SUVmax, and volume) were assessed alongside spatial resolution using full width at half maximum (FWHM) as a surrogate of motion‐induced blurring.

Results

Motion correction reduced myocardial FWHM on average (global mean decrease ∼0.6 mm), indicating improved spatial resolution and reduced motion‐related blurring. Threshold‐based segmentation yielded variable uptake and volume estimates due to spillover and saturation effects, whereas the fixed VOI approach demonstrated preserved quantitative uptake with minimal changes in SUVmean (< 1%). Motion correction reduced apparent myocardial volume, consistent with improved myocardial boundary definition. While variability in FWHM and SUV persisted due to redistribution of counts across respiratory bins, the overall effect reflected a balance between reduced motion blur and increased statistical noise.

Conclusions

Data‐driven respiratory motion correction improves myocardial spatial resolution and reduces regional differences in motion‐related blurring in [^{1 8}F]F‐AraG PET while preserving quantitative uptake measurements. Within a standard whole‐body PET workflow without dedicated cardiac gating, this approach provides a practical strategy to enhance image fidelity, particularly for tracers with subtle myocardial uptake.

1. INTRODUCTION

Positron emission tomography (PET) has been widely used for assessing myocardial viability and function in both clinical and research settings. ¹ [^{1 8}F]F‐AraG is a novel mitochondrial‐targeted PET tracer that selectively accumulates in activated T cells, offering the potential to detect immune‐related subclinical pathology in patients undergoing immunotherapy, and is currently being investigated for cardiac imaging applications. ² , ³

Accurate quantification of myocardial tracer uptake is compromised by respiratory and cardiac motion, which degrade spatial resolution, introduce regional variability, and increase measurement uncertainty. Traditional approaches such as respiratory gating rely on external hardware (e.g., pressure belts or spirometry), adding complexity and potential patient discomfort. Data‐driven gating (DDG) addresses this limitation by extracting motion signals directly from PET data, enabling correction without additional devices. ⁴ , ⁵ , ⁶ Prior studies have shown that respiratory motion alone can significantly alter apparent myocardial uptake in PET/CT, ⁴ , ⁷ underscoring the need for robust correction. Myocardial imaging presents additional challenges due to complex three‐dimensional motion from both respiration and contraction, combined with often subtle tracer uptake that is sensitive to partial‐volume effects and regional blurring. Consequently, spatial resolution is critical for accurate quantification, as even modest improvements can enhance regional quantification and the reliability of SUV measurements across myocardial segments. ⁸

OncoFreeze AI (Siemens Healthineers) is a vendor‐implemented, data‐driven respiratory motion correction framework for whole‐body PET, originally developed for oncologic imaging. It derives respiratory waveforms from PET coincidence data through analysis of spatiotemporal variations in lines of response (FlowMotion‐based DDG), without external hardware, and performs phase‐based binning followed by motion‐compensated reconstruction. ⁷ , ⁹ , ¹⁰ Although DDG is well established, prior studies have largely emphasized technical development and general image quality, with limited evaluation of its impact on myocardial quantification. This gap is particularly relevant for low‐uptake tracers such as [^{1 8}F]F‐AraG in standard whole‐body PET, where cardiac count statistics are limited and dedicated gating is often impractical. Therefore, this study investigates a practical clinical workflow and evaluates whether DDG improves myocardial spatial resolution, reduces regional differences in motion‐related blurring, and preserves quantitative uptake.

2. MATERIALS AND METHODS

2.1. Dataset

Paired non–motion‐corrected PET reconstructions (Regular) and motion‐corrected PET reconstructions processed with OncoFreeze AI were retrospectively analyzed from 15 healthy human subjects imaged at the UCSF China Basin Imaging Center. ³ All subjects were scanned on a Siemens Biograph Vision 600 system (Siemens Healthineers). Both Regular and OncoFreeze AI reconstructions were generated from the same raw PET data using vendor‐provided ordered subset expectation maximization with model‐based scatter correction and a 3D Gaussian post‐reconstruction filter (FWHM 3.5 mm). Quantitative image analysis was performed using OsiriX and AMIDE.

For the OncoFreeze AI reconstruction, respiratory waveforms were derived directly from PET list‐mode data without external monitoring devices. The algorithm combines Automated Landmarking and Parsing of Human Anatomy (ALPHA) technology for automatic identification of regions affected by respiratory motion with data‐driven respiratory gating. ¹¹ Motion‐frozen images are generated using an optimal‐flow algorithm that registers gated images to an optimal reference gate and recombines information from all respiratory phases into a single motion‐corrected image while preserving 100% of the acquired count statistics. This approach is particularly relevant in standard whole‐body PET imaging, where the effective acquisition time over the heart is approximately 2–3 min and count statistics are lower than those typically achieved in dedicated cardiac PET studies.

For spatial resolution analysis, left ventricular images were reoriented along the cardiac long axis to generate standard short‐ and long‐axis views. FWHM measurements were derived by fitting a Gaussian function to intensity line profiles extracted from myocardial segments defined by myocardial wall (septal, lateral, anterior) and ventricular level (apical, mid, basal).

2.2. Myocardial segmentation and VOI analysis

In the preliminary analysis, the myocardium was segmented using a threshold‐based approach with the “Grow Region (2D/3D Segmentation)” tool, applying a fixed standardized uptake value (SUV) threshold. An initial volume of interest (VOI) was manually defined using a 3D ball tool, after which the manual VOI was removed and automated thresholding was used to isolate myocardial tissue from surrounding structures. To improve reproducibility and reduce spillover from adjacent tissues, a refined analysis was performed using a small fixed‐size VOI (1 cm³), placed in the mid antero‐lateral wall of the left ventricle. This VOI was consistently replicated across paired reconstructions for each subject, ensuring spatial alignment and enabling extraction of SUVmean values with greater specificity and reduced sensitivity to boundary effects.

For threshold‐based segmentation, SUVmean, SUVmax, and myocardial volume were measured from the entire segmented myocardium in both OncoFreeze AI and Regular reconstructions. For the fixed VOI approach, only SUVmean was extracted from the predefined three‐dimensional VOI in the lateral wall.

Spatial resolution was quantified using full width at half maximum (FWHM) measured in matched myocardial segments reconstructed with OncoFreeze AI and Regular methods. Measurements were obtained from nine myocardial segments per subject, defined by myocardial wall (septal, lateral, anterior) and ventricular level (apical, mid, basal), yielding a fully crossed repeated‐measures dataset. Inferior wall segments were excluded due to substantial spillover from hepatic activity.

2.3. Statistical analysis

Statistical analyses were performed to assess (1) global differences in myocardial FWHM between reconstruction methods, (2) differences among myocardial walls and ventricular levels, and (3) whether the effects of motion correction varied by region. Data are reported as mean ± standard deviation (SD). Global differences between OncoFreeze AI and regular reconstructions were evaluated using two‐tailed paired t‐tests, with measurements paired within subjects (n = 15) and each myocardial segment serving as its own control; a two‐sided p value < 0.05 was considered statistically significant. ¹² Agreement between methods was assessed using Pearson correlation, while systematic bias and agreement were further evaluated using Bland–Altman analysis, with bias defined as the mean paired difference (OncoFreeze − Regular) and 95% limits of agreement calculated as bias ± 1.96 × SD of the differences.

Regional and method‐dependent effects were analyzed using repeated‐measures ANOVA (RM‐ANOVA) with fully within‐subject designs, including two‐factor models (Method × Wall; Method × Level) and a three‐factor model (Method × Wall × Level). Effect sizes were reported using partial η ². Paired differences in SUVmean and SUVmax between reconstructions were evaluated using two‐tailed paired t‐tests and p values reported to determine whether motion correction altered quantitative uptake measurements.

3. RESULTS

3.1. Quantitative uptake and myocardial volume

Across 15 subjects, myocardial SUVmean was similar between reconstructions (3.02 ± 0.29 vs. 3.01 ± 0.34), with a negligible percent change (p = 0.72), indicating no significant difference in global quantitative uptake with OncoFreeze AI. In contrast, SUVmax increased significantly from 5.26 ± 0.62 to 5.56 ± 0.69 (p = 0.0002), consistent with improved peak signal recovery due to reduced blurring. Mean myocardial volume decreased from 200.9 ± 50.8 mL to 172.4 ± 40.0 mL (p = 0.0004), likely reflecting improved boundary delineation.

Using a fixed VOI approach, SUVmean increased slightly from 2.72 ± 0.18 to 2.74 ± 0.16 (p = 0.009). Although modest, this change reached significance due to low inter‐subject variability. Overall, these findings suggest that motion correction preserves quantitative uptake while improving image sharpness, with consistent measurements indicating reduced spillover from adjacent tissues without materially altering myocardial uptake.

Figure 1 presents a representative qualitative comparison of [^{1 8}F]F‐AraG images between motion‐corrected and non–motion‐corrected reconstructions. Although differences are subtle, the myocardium appears darker and more sharply defined after motion correction, consistent with reduced respiratory blurring and improved boundary delineation. The adjacent liver (circled) also shows sharper contours and higher contrast.

FIGURE 1.

Representative [^{1 8}F]F‐AraG PET images demonstrating the effect of motion correction. Motion‐corrected reconstructions (a) using OncoFreeze AI show reduced blurring and sharper myocardial boundaries compared with non–motion‐corrected images (b). The arrow highlights the myocardium. The circled region indicates the adjacent liver, which appears more sharply contoured and higher in contrast following motion correction. Coronal images shown at the bottom further demonstrate improved myocardial delineation with motion correction.

3.2. Spatial resolution (FWHM)

Bland–Altman analysis of FWHM measurements (Figure 2a) demonstrated a mean bias of −0.65 mm, indicating systematically lower FWHM values with OncoFreeze AI. A significant proportional bias was observed, with larger reductions in FWHM occurring at higher mean FWHM values (slope = −0.23 mm/mm, p < 0.001). This finding indicates that motion correction provides greater benefit in segments with greater apparent blurring, consistent with the expectation that regions affected by more severe motion‐related degradation have greater potential for improvement following correction.

FIGURE 2.

(a) Bland–Altman analysis of myocardial FWHM comparing OncoFreeze AI and regular reconstruction. The solid line denotes the mean bias (−0.65 mm), and the dashed lines indicate the 95% limits of agreement (−2.60 to 1.30 mm). A significant proportional bias was observed (slope = −0.23 mm/mm, p < 0.001), indicating greater FWHM reductions at higher mean FWHM values. (b) Comparison of myocardial FWHM between motion‐corrected and non–motion‐corrected reconstructions. Bars represent mean ± SD across 15 subjects for septal, lateral, and anterior walls at apical, mid‐ventricular, and basal levels, as well as the global average.

Regionally (Figure 2b), RM‐ANOVA showed significant main effects of method (p < 0.001, partial η ² = 0.48) and wall (p < 0.001, partial η ² = 0.22), as well as a significant interaction (p = 0.021, partial η ² = 0.08). Post‐hoc analyses, averaged across ventricular levels, demonstrated significant reductions in FWHM with OncoFreeze AI in the lateral (12.36 ± 1.61 vs. 12.81 ± 1.81 mm, p < 0.001) and anterior walls (12.54 ± 1.49 vs. 14.01 ± 1.69 mm, p < 0.001), whereas the septal wall showed no significant change (11.47 ± 1.27 vs. 11.52 ± 1.22 mm, p = 0.58), indicating that the improvement was driven primarily by the anterior and lateral regions. FWHM reductions were also significant across ventricular levels, with the greatest reduction observed in basal segments (12.33 ± 1.71 vs. 13.13 ± 2.08 mm, p < 0.001), followed by mid‐ventricular (11.75 ± 1.54 vs. 12.45 ± 1.92 mm, p < 0.001) and apical segments (12.30 ± 1.25 vs. 12.76 ± 1.45 mm, p = 0.003).

In the full three‐way RM‐ANOVA, the main effect of method remained significant (p < 0.001, partial η ² = 0.51), together with significant wall and level effects, indicating that the magnitude of FWHM reduction varied by myocardial location, with the greatest improvements observed in motion‐prone regions.

4. DISCUSSION

This study extends DDG using OncoFreeze AI—previously applied primarily in oncologic PET—to the evaluation of subtle myocardial tracer uptake in routine whole‐body PET imaging. Using [^{1 8}F]F‐AraG as a representative tracer, we show that device‐free motion correction improves spatial resolution, reduces regional differences in motion‐related blurring, and preserves quantitative robustness under clinically realistic conditions. Unlike prior cardiac PET studies relying on dedicated acquisitions and comprehensive gating strategies, this approach operates without additional hardware or ECG gating, aligning with standard clinical workflows. These findings support a practical and generalizable framework for enhancing myocardial PET image fidelity, particularly for tracers with low or heterogeneous uptake. We note that OncoFreeze AI is a general‐purpose respiratory motion correction tool designed for whole‐body PET and is distinct from dedicated cardiac motion‐correction solutions such as CardioFreeze (Siemens Healthineers), which typically incorporate cardiac gating and are optimized for longer, cardiac‐focused acquisitions. Accordingly, the present results should be interpreted within the context of a standard whole‐body PET workflow with limited cardiac acquisition time, rather than as a direct surrogate for fully optimized cardiac PET protocols.

Motion correction reduced myocardial blurring on average across subjects and segments, underscoring the impact of respiratory motion on cardiac PET resolution. Strong concordance between corrected and uncorrected measurements indicates that motion correction preserves the underlying signal while improving spatial resolution. A significant proportional bias was observed, with larger reductions in FWHM occurring at higher mean FWHM values, suggesting greater benefit in segments exhibiting greater apparent blurring. Although both reconstructions originated from the same list‐mode data, DDG redistributes events into respiratory bins, reducing counts per frame prior to recombination and introducing a trade‐off between motion reduction and increased noise. This likely contributes to the observed spread in SUV and FWHM measurements (Figure 2). While most segments improved, a subset showed increased FWHM without clustering by region or subject, suggesting non‐systematic effects related to noise, count variability, and limitations in respiratory signal extraction.

Despite persistent wall‐ and level‐based differences, motion correction reduced regional disparities in spatial resolution across the myocardium. Septal segments exhibited the least blurring, lateral segments intermediate values, and anterior segments the greatest—consistent with known motion gradients—while basal segments remained most susceptible due to proximity to the diaphragm and atrioventricular plane. Improvement was most pronounced in these motion‐prone regions, reducing inter‐wall and apical‐to‐basal differences in spatial resolution. Such reductions in regional disparities may facilitate more robust regional quantification, particularly in longitudinal studies.

Quantitative uptake analyses further support a stabilizing effect of motion correction. VOI‐based measurements demonstrated preserved or mildly increased tracer uptake after correction, consistent with improved signal localization and reduced partial‐volume effects rather than artificial inflation. The accompanying reduction in apparent myocardial volume is consistent with improved myocardial boundary definition following motion correction.

Several limitations should be considered. First, this study was conducted in a relatively small cohort of healthy subjects (n = 15), which may limit generalizability to patients with cardiac disease or more complex motion patterns. Motion correction was evaluated using a single data‐driven approach, and the findings may not generalize to other methods. Spatial resolution was assessed indirectly using FWHM without a motion‐free reference standard, and simplified segmentation may underestimate benefits in more motion‐prone regions while limiting assessment of global myocardial heterogeneity. Metrics derived from established cardiac PET analysis software (e.g., Emory Cardiac Toolbox, 4DM, and QPS/QGS) were not evaluated, as this study focused on fundamental image quality metrics within a whole‐body PET setting. Finally, the clinical implications and tracer‐specific effects of [^{1 8}F]F‐AraG require validation in larger, disease‐specific cohorts, and integration with functional and polar map–based analyses remain an important direction for future work.

5. CONCLUSION

OncoFreeze AI improves myocardial spatial resolution and reduces regional differences in motion‐related blurring, while only modestly affecting SUV measurements. These findings highlight the value of data‐driven respiratory motion correction as a practical approach in whole‐body PET, particularly for low‐uptake tracers and precision‐focused studies.

CONFLICT OF INTEREST STATEMENT

JL and HC are employees of CellSight Technologies, Inc., the developer of [^{1 8}F]F‐AraG. All other authors declare no competing interests.

DATA AVAILABILITY STATEMENT

The data supporting the findings of this study are available from the corresponding author upon reasonable request.