A rapid total-body PET imaging approach for pediatric patients using non-attenuation-corrected PET scans Authors: Jingxi Hu, Yuqian Huang, Qiyang Zhang, Yingying Hu, Yumo Zhao, Xu Zhang, Wei Fan, Hairong Zheng, Dong Liang, Zhanli Hu, Changhui Jiang, Xiangjian He, Na Zhang Published: 2026 ABSTRACT BackgroundPediatric lymphoma patients undergo multiple 18F-FDG PET/CT examinations for staging and response assessment, raising concerns about cumulative radiation dose, particularly from the CT component. We propose SnapPET, a CT-sparing deep learning–based framework that uses 2D ultra-short non-attenuation-corrected (NAC) PET (Maximum intensity projection) MIP images to generate 2D high-quality PET MIP images for rapid screening and triage.MethodsTwo hundred pediatric patients with FDG-avid lymphoma imaged on a total-body PET/CT scanner were retrospectively analyzed. From 10-min list-mode data, non-attenuation-corrected (NAC) PET images at 3–300 s were retrospectively reconstructed, and 600-s attenuation-corrected (AC) PET served as the reference. Whole-body coronal maximum-intensity projections (MIPs) were converted into 2D images and divided into 64 × 64 patches. For each acquisition duration, an independent latent diffusion–based autoencoder was trained to map the corresponding short-duration NAC PET MIPs to AC-like coronal PET MIPs. The intended clinical application was rapid pre-CT triage, with primary emphasis on the ultra-short 3-s setting, while longer durations were analyzed comparatively across count levels. Image quality was evaluated using PSNR, SSIM, and NMSE, as well as lesion ROI–based SUVmax and SUVmean. Three benchmark models (U-Net, GAN, and DDPM) were trained under identical settings for comparison. In addition, two nuclear medicine physicians assessed image quality using 5-point Likert scales, and lesion-level ROC and ALROC analyses were performed.ResultsSnapPET substantially improved image quality over NAC input for all acquisition times. For 3-s scans, PSNR increased from 28.6 ± 4.3 to 35.3 ± 6.0 and SSIM from 0.85 ± 0.09 to 0.98 ± 0.02, while NMSE decreased from 0.23 ± 0.18 to 0.03 ± 0.01 (all p < 0.001). SUV metrics in lesion ROIs showed high correlation with reference AC PET, with minimal bias. Compared with UNet, GAN, and DDPM, SnapPET achieved the best quantitative metrics and the highest lesion-level ROC and ALROC curves, indicating superior detectability and localization. Reader scores for noise suppression, artifact reduction, contrast retention, and lesion discrimination were consistently higher for SnapPET images, particularly for 3- and 15-s acquisitions.ConclusionSnapPET enables rapid pediatric PET imaging by generating high-quality, AC-like 2D images directly from ultra-short NAC acquisitions. The method improves image quality, preserves quantitative accuracy, and enhances lesion detectability compared with both NAC input and competing deep-learning baselines. This CT-sparing screening and triage approach has the potential to reduce reliance on attenuation-correction CT and support personalized, lower-dose CT strategies in pediatric lymphoma care.Supplementary InformationThe online version contains supplementary material available at 10.1186/s40658-026-00889-x. Introduction Hodgkin and non-Hodgkin lymphomas together represent a major proportion of pediatric malignancies, with Hodgkin lymphoma accounting for approximately 5% of childhood cancers and non-Hodgkin lymphoma ranking as the fourth most common malignancy in children and adolescents [1–5]. Owing to advances in chemoradiotherapy and risk-adapted protocols, the prognosis for these patients has markedly improved, with 5-year survival rates exceeding 98% for Hodgkin lymphoma and 80% for non-Hodgkin lymphoma [2, 6, 7]. However, the management of pediatric lymphoma typically requires multiple 18F-FDG PET/CT examinations throughout the clinical course—commonly at initial staging, interim response assessment after 2–3 therapy cycles, end-of-therapy evaluation, and follow-up surveillance for relapse [8, 9]. This repeated imaging schedule, while essential for guiding treatment decisions within standardized response criteria such as the Lugano classification, results in substantial cumulative radiation exposure over a relatively short timeframe, i.e., within weeks to months during a treatment course [10, 11]. A standard positron emission tomography/computed tomography (PET/CT) examination typically begins with a low-dose 2D scout/topogram (localization view) for patient positioning and planning the subsequent CT acquisition; because it is a projection rather than a tomographic scan, its radiation dose is very low. After this scout image, one tomographic CT acquisition is performed—either a low-dose CT primarily for localization and attenuation correction (AC) or a diagnostic CT, which also provides AC information, but not both within a single examination. In pediatric patients, these CT options contribute differently to total radiation dose. Typically, the effective dose from an AC/localization CT is about 0.5–1.0 mSv, while that from a diagnostic CT range from 5–10 mSv, depending on body region and scanning parameters [12, 13]. Combined with the radiotracer dose (commonly 3–5 MBq/kg in adults but reduced to approximately 1–2 MBq/kg in children to mitigate radiation risk [14–16]), the overall burden of repeated PET/CT scans becomes a critical concern in long-term survivors. Minimizing radiation exposure, especially from the CT component, is therefore an important clinical goal in pediatric oncology imaging [12, 13]. For pediatric lymphoma, diagnostic-quality CT is often required in addition to 18F-FDG PET to fully characterize nodal and extranodal involvement, assess lesion size and morphology, evaluate lung and mediastinal parenchymal changes, and delineate the relationship to adjacent organs and vessels. These anatomical details are important for staging, risk stratification, treatment planning (e.g., radiotherapy fields), and differential diagnosis (e.g., distinguishing infection/inflammation from tumor). Recent advances in total-body PET/CT systems have enabled substantially higher sensitivity and faster acquisitions, allowing significant reduction in injected activity and/or acquisition time without compromising diagnostic quality [17]. In parallel, several AI-based methods have focused on denoising or reconstructing low-count PET data to enhance image quality, using deep learning or model-based approaches for low-dose and ultra-low-dose imaging [18–20]. Other emerging applications of AI in PET/CT include automated detection, diagnosis support, and task-specific reconstruction in oncologic and neurologic imaging [21–23]. However, most of these studies primarily target PET-only improvements, while the CT component remains the dominant source of radiation in pediatric PET/CT workflows, and existing strategies rarely address workflow-level optimization—such as selectively omitting ACCT or personalizing diagnostic CT protocols according to disease presentation [24, 25]. Thus, a clinically integrated, AI-assisted approach is needed to streamline the imaging pathway while maintaining diagnostic confidence[26–28]. To address this gap, we propose SnapPET, a deep learning–based rapid screening and triage framework specifically designed for pediatric lymphoma patients. Methodologically, SnapPET is an image-to-image reconstruction model: it takes ultra-short, 3-s non-attenuation-corrected (NAC) 2D PET as input and predicts a high-quality 2D PET image that approximates the corresponding attenuation-corrected (AC) PET. In other words, SnapPET functions as a CT-sparing, pre-CT triage image-enhancement tool that generates an AC-like PET MIP from ultra-short NAC inputs, aiming to provide sufficient image quality for rapid preliminary assessment of 18F-FDG–avid lesions, rather than to replace full diagnostic AC PET/CT [8, 9]. The intended clinical use of SnapPET is illustrated in Fig. 1. During the initial scan, only a 3-s NAC 2D coronal PET MIP acquisition is performed, which is converted by SnapPET into an enhanced AC-like 2D coronal whole-body MIP image. If no abnormal 18F-FDG uptake is observed on this reconstructed image, the patient proceeds only to a low-dose CT for basic anatomical documentation. If 18F-FDG-avid lesions are detected, SnapPET enables a personalized CT strategy in which the CT component is performed as a single continuous CT scan rather than two separate acquisitions. A low-dose CT protocol is used for regions without 18F-FDG-avid uptake to provide attenuation correction and basic anatomical localization, while bed positions containing 18F-FDG-avid lesions on SnapPET can be scanned with a diagnostic CT protocol (higher tube current and finer slice thickness, with optional contrast if clinically indicated). Modern PET/CT systems support such bed-specific parameter adjustments, enabling this region-adapted CT strategy within one integrated scan so that high-resolution anatomical information is acquired only where it is clinically needed. In this way, SnapPET is primarily designed to support rapid triage and workflow optimization, using improved but still screening-oriented image quality to safely reduce unnecessary diagnostic CT exposure while maintaining the accuracy required for lymphoma management, in line with the ALARA (As Low As Reasonably Achievable) principle. All PET/CT scans were acquired on an Imaging Healthcare uEXPLORER total-body PET/CT scanner.Fig. 1Schematic of the SnapPET triage workflow. A NAC 2D coronal PET MIP scan is first acquired and enhanced by the SnapPET network into an AC-like 2D coronal MIP image. If no abnormal 18F-FDG uptake is observed, the patient undergoes a single low-dose CT for anatomical localization. If 18F-FDG-avid lesions are detected, a single continuous CT scan is performed with bed-specific protocols: diagnostic CT parameters for 18F-FDG-avid regions and low-dose CT parameters for the remaining body, thereby reducing overall radiation exposure Schematic of the SnapPET triage workflow. A NAC 2D coronal PET MIP scan is first acquired and enhanced by the SnapPET network into an AC-like 2D coronal MIP image. If no abnormal 18F-FDG uptake is observed, the patient undergoes a single low-dose CT for anatomical localization. If 18F-FDG-avid lesions are detected, a single continuous CT scan is performed with bed-specific protocols: diagnostic CT parameters for 18F-FDG-avid regions and low-dose CT parameters for the remaining body, thereby reducing overall radiation exposure In this study, we propose SnapPET, a CT-sparing deep learning–based triage framework for pediatric oncology. SnapPET reconstructs AC-like 2D coronal whole-body MIP images from 3-s NAC acquisitions and is evaluated both for its ability to approximate standard AC PET and for its impact on lesion detectability in expert reader studies. Materials and methods Participants and image acquisition As shown in Table 1, the employed data were collected from Sun Yat-sen University Cancer Center, encompassing a total of 200 pediatric 18F-FDG-avid lymphoma patients (135 males and 65 females) with an age distribution of 11.22 ± 6.39 years and ranging from 3 to 18 years. Approximately 55 to 60 min following the administration of the full-dose of 18F-FDG tracer (0.1 mCi/kg), raw data were collected over 10 min. The data acquisition period spanned from August 2020 to October 2020.Table 1Information on the participants and injection doses used in this studyAge (years old) 1 ~ 663 7 ~ 1369 14 ~ 1968Sex Male135 Female65Weight* (kg)29.02 ± 18.43Height (cm)125.41 ± 31.13BMI (kg/m2)16.51 ± 3.26Injected dose (mbq)108.5 ± 67.6Dose per kg (mbq/kg)3.77 ± 2.28 Table 1Information on the participants and injection doses used in this studyAge (years old) 1 ~ 663 7 ~ 1369 14 ~ 1968Sex Male135 Female65Weight* (kg)29.02 ± 18.43Height (cm)125.41 ± 31.13BMI (kg/m2)16.51 ± 3.26Injected dose (mbq)108.5 ± 67.6Dose per kg (mbq/kg)3.77 ± 2.28 Information on the participants and injection doses used in this study For each patient, we retrospectively reconstructed non-attenuation-corrected (NAC) PET images at multiple durations (3 s, 6 s, 15 s, 20 s, 30 s, 40 s, 60 s, 150 s, and 300 s) from the same 600-s list-mode acquisition. For each target duration, the first contiguous N-second segment was extracted from the list-mode stream according to event timestamps (one replicate per duration), and the truncated events were rebinned into sinograms for reconstruction. Normalization correction and randoms correction were applied to both the short-duration NAC reconstructions and the 600-s reference reconstruction using the scanner’s standard vendor-provided reconstruction pipeline. For the NAC images, attenuation correction was disabled and scatter correction was not applied. In contrast, the 600-s attenuation-corrected (AC) reference reconstruction included attenuation correction and scatter correction in addition to normalization and randoms correction. The same reconstruction software and parameter settings were used across durations unless otherwise specified. Generate training data Given the objective of expediting lesion detection from NAC low-dose imaging, our model operates on 2D PET representations. Volumetric total-body PET scans are first converted into \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$672\times 256$$\end{document} 2D images using coronal maximum intensity projection (MIP), which projects the maximum voxel value along each ray to summarize 3D uptake in a single 2D view. This routinely used approach allows rapid, holistic assessment of total-body lesion burden and continuity without scrolling through individual slices. Although the MIP images are of moderate size, a patch-based training strategy (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$64\times 64$$\end{document}) is employed to augment the dataset and prevent overfitting. Randomly sampled patches increase the diversity of local 18F-FDG uptake patterns, enabling the model to better learn fine-grained features of small lesions while maintaining computational efficiency; the enhanced images are subsequently reconstructed from overlapping patches to preserve global spatial continuity. The resulting 2D MIP images are used as inputs to the deep learning model, with the corresponding 600-s 2D AC PET MIPs serving as ground truth [29, 30].The dataset, acquired at different scanning durations, is randomly split into training and testing sets in an 8:2 ratio, yielding 160 patients for training and 40 for testing. To alleviate the limitation of a relatively small training cohort and improve robustness, we additionally perform tenfold cross-validation. Model framework and training As shown in Fig. 2, SnapPET is a latent diffusion autoencoder operating on 2D whole-body coronal MIP PET images [31]. We adopted a latent diffusion–based autoencoder rather than a stand-alone image-to-image U-Net because ultra-short NAC PET MIPs are severely degraded by count-limited noise and attenuation-related intensity distortion, making direct restoration in the image domain challenging [32, 33]. The autoencoder first maps the NAC input and the corresponding AC target into a compact latent space that preserves the major uptake pattern while reducing redundancy and high-frequency noise, thereby making the restoration task better conditioned than direct pixel-space prediction. Conditional diffusion is then used to iteratively refine the degraded latent representation toward the AC-like target distribution. Compared with a one-step deterministic mapping, this progressive denoising process is better suited to suppressing mottled noise while preserving lesion-to-background contrast and the overall physiological uptake distribution. Performing diffusion in the latent space also improves computational efficiency and training stability compared with full image-domain diffusion. A direct comparison between the original and revised architectures is provided in the Supplementary Material. A low-quality 3D PET reconstruction is first converted to a 2D NAC MIP, which is encoded by an autoencoder \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{E}}(\cdot )$$\end{document} into a compact latent representation \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{z}}}_{{{y}}}={{E}}({{y}})$$\end{document}. The corresponding high-quality reference AC MIP \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x$$\end{document} is encoded as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{z}}}_{{{x}}}={{E}}({{x}})$$\end{document}. In the latent space, a conditional denoiser \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{\epsilon}}}_{{{\theta}}}(\cdot )$$\end{document}(attention U-Net) performs iterative denoising conditioned on \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{z}}}_{{{y}}}$$\end{document}, and the denoised latent is decoded by \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{D}}(\cdot )$$\end{document} to obtain the output \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\widehat{{{x}}}={{D}}({\widehat{{{z}}}}_{0})$$\end{document} [34]. Fig. 2Overview of SnapPET. A low-quality 3D NAC PET reconstruction is converted to a 2D whole-body coronal maximum-intensity-projection (MIP) image and encoded by a pretrained autoencoder into a latent representation. A conditional latent diffusion denoiser (attention U-Net) iteratively refines the latent conditioned on the low-quality input, and the decoded output yields the restored 2D AC MIP. For training, we use NAC inputs acquired at multiple scan durations (3, 6, 15, 20, 30, 40, 60, 150, and 300 s), with the 600-s scan reconstructed with attenuation correction serving as the reference Overview of SnapPET. A low-quality 3D NAC PET reconstruction is converted to a 2D whole-body coronal maximum-intensity-projection (MIP) image and encoded by a pretrained autoencoder into a latent representation. A conditional latent diffusion denoiser (attention U-Net) iteratively refines the latent conditioned on the low-quality input, and the decoded output yields the restored 2D AC MIP. For training, we use NAC inputs acquired at multiple scan durations (3, 6, 15, 20, 30, 40, 60, 150, and 300 s), with the 600-s scan reconstructed with attenuation correction serving as the reference Latent diffusion objective During training, we sample a timestep \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{t}}\in \{1,\dots ,{{T}}\}$$\end{document} and perturb the target latent \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{z}}}_{{{x}}}$$\end{document} using a fixed forward noising process:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$z_{t} = \sqrt {\bar{\alpha }_{t} } \,z_{x} + \sqrt {1 - \bar{\alpha }_{t} }\,\, \epsilon, \epsilon\, \sim {\mathcal{N}}\left( {0,I} \right),$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\bar{\alpha }}_{t}$$\end{document} is defined by the noise schedule. The denoiser predicts the injected noise \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\epsilon$$\end{document} from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left({z}_{t},t,{z}_{y}\right)$$\end{document} and is optimized with the standard noise-prediction loss\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathcal{L}}_{diff}={\mathbb{E}}_{t,\epsilon }\left[{\parallel \epsilon -{\epsilon}_{\theta }({z}_{t},t,{z}_{y})\parallel }_{2}^{2}\right].$$\end{document} After reverse denoising, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\widehat{z}}_{0}$$\end{document} is decoded to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\widehat{x}$$\end{document}. To encourage image-space fidelity, we additionally use a pixel-wise reconstruction loss and a perceptual loss:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal{L}={\mathcal{L}}_{diff}+{\lambda}_{rec}\parallel \widehat{x}-x{\parallel}_{1}+{\lambda}_{perc}{\mathcal{L}}_{perc}(\widehat{x},x),$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathcal{L}}_{perc}$$\end{document} is computed from VGG-19 feature maps (conv3_3 and conv4_3). The diffusion schedule (variance parameters and number of timesteps \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$T$$\end{document}) is fixed; only the parameters of the latent denoising U-Net are optimized (the autoencoder is kept fixed after pretraining). Since the autoencoder is frozen after pretraining, both the reconstruction and perceptual losses are backpropagated through the decoder and are used to update only the denoising U-Net parameters. In all experiments, we set \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\lambda}_{\mathrm{r}\mathrm{e}\mathrm{c}}=0.1$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\lambda}_{\mathrm{p}\mathrm{e}\mathrm{r}\mathrm{c}}=0.01.$$\end{document} Implementation details For training, 2D MIP images are divided into \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$64\times 64$$\end{document} patches, which are randomly sampled to form batches [35, 36]. The model is trained for 300 epochs with a batch size of 25 using Adam with an initial learning rate of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1\times {10}^{-4}$$\end{document} [37–39]. Convolutional weights are initialized with He initialization, gradient norms are clipped to stabilize training, and the entire pipeline is implemented in PyTorch on an NVIDIA A6000 GPU. To analyze performance across count levels, we train independent models for each acquisition duration (3, 6, 15, …, 300 s) and evaluate each model on the corresponding duration [40, 41]. The intended rapid triage setting emphasizes ultra-short inputs (primarily 3 s, and also 6/15 s), while longer durations are reported as a comparative analysis. Evaluation metrics We assessed image quality using both objective and subjective metrics. The objective metrics included mean squared error (MSE), normalized mean squared error (NMSE), peak signal-to-noise ratio (PSNR), and structural similarity index (SSIM), all computed pixel-wise between the reconstructed images and the corresponding reference images [29, 42–44]. MSE and NMSE quantify the overall squared intensity differences (with NMSE normalized by the reference image energy), PSNR reflects the reconstruction fidelity relative to the dynamic range of the image, and SSIM evaluates structural similarity by jointly considering local luminance, contrast, and texture. Together, these metrics provide complementary numerical measures of how closely the generated images approximate the ground-truth PET images and help quantify the effectiveness of the proposed method. In addition to these image-based metrics, we also evaluated quantitative accuracy using the standardized uptake value (SUV), defined as:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{SUV}} = \frac{{C_{{PET}} \left( t \right)}}{{Dose/Weight}} .$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{C}}}_{PET}$$\end{document} (kBq/mL) refers to the radioactivity per milliliter in a specific tissue or lesion. The dose (MBq) is the total amount of radiotracer received by the patient. “Weight (kg)” is the patient weight. In this study, we selected a rectangular region of interest (ROI) on total-body PET images from 40 patients to encompass pathological lesions. A single rectangular ROI was manually delineated for each patient by an experienced nuclear medicine physician, ensuring the inclusion of the entire lesion and minimizing inclusion of adjacent normal tissues. The ROI boundaries were applied uniformly across different scan time groups for consistent analysis. Using Python 3.9, we calculated semi-quantitative metrics, including SUVmax, SUVmean, and standard deviation(std) within the defined ROI. The calculations were performed using the following Python packages: pydicom version 3.0.1, pandas version 2.2.3, numpy version 2.1.3, and torch version 2.0.1. This approach allowed for a comprehensive and standardized assessment of lesion uptake across various scan protocols, facilitating direct comparison of PET tracer uptake among different patient populations. Clinical readings Clinical image quality and lesion detectability were evaluated by two board-certified nuclear medicine physicians with at least three years of experience in pediatric PET/CT. All reconstructed PET images (NAC input, UNet, GAN, DDPM, and SnapPET) were anonymized and randomly presented, with readers blinded to reconstruction method and acquisition time and reviewing them alongside the corresponding full-quality reference scans.For image-quality assessment, each image was scored on a 5-point Likert scale for artifact reduction, noise suppression, contrast retention, and lesion discrimination, where 1= “Uninterpretable”, 2 = “Poor”, 3 = “Adequate”, 4 = “Good”, and 5 = “Excellent”. For lesion analysis, cases were classified as positive or negative according to the Lugano criteria, with lesions showing 18F-FDG uptake exceeding mediastinal background (Deauville score ≥ 3) labeled as positive [10, 11]. The two readers marked 18F-FDG-avid lymphoma lesions by indicating lesion location and a 5-point confidence score for malignancy. For ROC analysis, these lesion-level confidence scores were used as decision variables to count true and false positives and generate ROC curves. For ALROC, a detection was considered a true positive only if the marked location fell within the reference lesion volume on the 600-s AC PET, thereby incorporating both detection and localization; ROC and ALROC curves were computed for each reader and reconstruction method and then averaged for comparison. Statistical analysis The results are reported as means ± std or frequencies and percentages. To compare differences among the PSNRs, SSIMs and tumor SUVs of the PET dataset images acquired at different time points, we employ paired t tests and a Bland‒Altman analysis. All the statistical analyses are performed using the Seaborn (version 0.11.0) and SciPy (version 1.6.0) Python libraries. A probability value of < 0.05 is considered the level of statistical significance. Results Image quality assessment metrics Based on Fig. 3, the generated images demonstrate an improvement over the original low-quality input images by reducing noise and preserving key features. In both coronal and sagittal views, the generated images more closely resemble the high-quality ground truth, capturing finer details and showing greater clarity in high-uptake regions. This suggests the model's potential to enhance image quality by reducing noise while retaining structural integrity.Fig. 3The generated images for 3 s, 15 s, and 60 s, along with their corresponding input images, use the 600-s AC image as the ground truth The generated images for 3 s, 15 s, and 60 s, along with their corresponding input images, use the 600-s AC image as the ground truth Table 2 presents comparing the generated and input MIP images using metrics such as PSNR, SSIM, and NMSE. The average PSNR, NMSE, and SSIM values indicate that the generated images yield strong performance across all three-evaluation metrics, confirming the reliability and quality of the generated images. A paired t-test shows significant differences (p < 0.05) between the input and output data. When comparing model outputs at different acquisition durations, the quantitative image quality metrics (PSNR, SSIM, and NMSE) show the most substantial improvements at ultra-short acquisition times (3 s, 6 s, and 15 s). In these cases, the model significantly enhances visual clarity, especially in high-uptake regions that were poorly distinguishable in the original NAC 2D coronal PET MIP images.Table 2Quantitative results obtained for the images in the test datasets. The data are presented as mean ± std valuesTimeOriginalProposedPSNRSSIMNMSEPSNRSSIMNMSE3-s28.6 ± 4.270.851 ± 0.0870.225 ± 0.1836.0 ± 5.570.981 ± 0.0180.029 ± 0.0076-s31.9 ± 4.390.875 ± 0.074 0.096 ± 0.0737.1 ± 5.220.982 ± 0.0180.020 ± 0.00715-s35.8 ± 5.860.916 ± 0.0450.034 ± 0.0238.4 ± 4.130.987 ± 0.0120.017 ± 0.00620-s36.8 ± 5.820.951 ± 0.0360.026 ± 0.01838.9 ± 3.900.986 ± 0.0110.016 ± 0.00630-s38.0 ± 5.910.961 ± 0.0340.020 ± 0.01439.2 ± 3.780.988 ± 0.0110.014 ± 0.00640-s38.9 ± 5.800.973 ± 0.0230.015 ± 0.01039.7 ± 3.750.990 ± 0.0100.013 ± 0.00560-s39.0 ± 5.700.981 ± 0.0180.012 ± 0.00739.9 ± 3.610.991 ± 0.0080.011 ± 0.0.004150-s40.7 ± 5.720.991 ± 0.0090.006 ± 0.00441.2 ± 3.590.992 ± 0.0070.010 ± 0.0.003300-s40.8 ± 4.840.991 ± 0.0090.005 ± 0.00341.9 ± 3.600.992 ± 0.0060.009 ± 0.003 Table 2Quantitative results obtained for the images in the test datasets. The data are presented as mean ± std valuesTimeOriginalProposedPSNRSSIMNMSEPSNRSSIMNMSE3-s28.6 ± 4.270.851 ± 0.0870.225 ± 0.1836.0 ± 5.570.981 ± 0.0180.029 ± 0.0076-s31.9 ± 4.390.875 ± 0.074 0.096 ± 0.0737.1 ± 5.220.982 ± 0.0180.020 ± 0.00715-s35.8 ± 5.860.916 ± 0.0450.034 ± 0.0238.4 ± 4.130.987 ± 0.0120.017 ± 0.00620-s36.8 ± 5.820.951 ± 0.0360.026 ± 0.01838.9 ± 3.900.986 ± 0.0110.016 ± 0.00630-s38.0 ± 5.910.961 ± 0.0340.020 ± 0.01439.2 ± 3.780.988 ± 0.0110.014 ± 0.00640-s38.9 ± 5.800.973 ± 0.0230.015 ± 0.01039.7 ± 3.750.990 ± 0.0100.013 ± 0.00560-s39.0 ± 5.700.981 ± 0.0180.012 ± 0.00739.9 ± 3.610.991 ± 0.0080.011 ± 0.0.004150-s40.7 ± 5.720.991 ± 0.0090.006 ± 0.00441.2 ± 3.590.992 ± 0.0070.010 ± 0.0.003300-s40.8 ± 4.840.991 ± 0.0090.005 ± 0.00341.9 ± 3.600.992 ± 0.0060.009 ± 0.003 Quantitative results obtained for the images in the test datasets. The data are presented as mean ± std values Qualitative comparison with additional deep-learning baselines Figure 4 presents qualitative results for two representative pediatric patients, comparing the low-dose 3-s NAC 2D coronal MIP input, UNet [45], GAN [46], DDPM [47]and the proposed SnapPET reconstruction against the high-quality 10-min AC 2D coronal MIP PET reference [5]. The UNet model reduces noise but tends to oversmooth the liver and mediastinal regions, with partial loss of small 18F-FDG-avid lesions. The GAN-based method produces sharper uptake patterns but occasionally introduces patchy artificial textures, particularly in the skeletal system. The DDPM model provides better noise suppression than UNet and GAN yet still exhibits residual mottled noise and reduced lesion–background contrast in the abdomen. In contrast, SnapPET more faithfully preserves both lesion uptake and normal physiological distribution, while markedly suppressing background noise, yielding the closest visual agreement with the 10-min AC PET images in both cases.Fig. 4Representative whole-body 2D PET MIP images from two pediatric cases, comparing the low-dose 3-s NAC input, UNet, GAN, DDPM, and the proposed SnapPET method against the 10-min high-quality AC reference Representative whole-body 2D PET MIP images from two pediatric cases, comparing the low-dose 3-s NAC input, UNet, GAN, DDPM, and the proposed SnapPET method against the 10-min high-quality AC reference ROI‑based assessment In addition to overall image analysis, we performed quantitative assessments on regions of interest (ROIs) with higher SUV values. Figure 5 shows the results using a violin plot. As expected, the SUV values in the LQ PET images are lower than in the AC PET images. This difference occurs because attenuation correction adjusts for radiation absorption and scattering, providing a more accurate measurement of the tissue’s uptake of the radiotracer. Without this correction, LQ PET images may underestimate radiotracer uptake. We also found that the distribution of SUV values in our generated PET images closely matches the ground truth. The median and quartile ranges of our images are nearly identical to those of the true AC PET images.Fig. 5The violin plots for our generated PET images within the different sites. “LQ” denotes the 3 s low quality NAC 2D coronal PET MIP images. “GT” denotes the high quality 600 s AC 2D coronal PET MIP images The violin plots for our generated PET images within the different sites. “LQ” denotes the 3 s low quality NAC 2D coronal PET MIP images. “GT” denotes the high quality 600 s AC 2D coronal PET MIP images To evaluate how well our generated PET images correlate with the ground truth in the ROIs, we used the Pearson correlation coefficient (PCC). Additionally, t-tests were conducted. Our generated PET images showed a strong correlation with the ground truth, yielding high PCC values. For example, for ROIs in the head, the PCC reached 0.90 (p < 0.05), and the SUV distributions and medians were consistent with the ground truth. Clinical readings To rigorously assess the clinical utility of the generated images, we conducted a reader study using subjective visual quality metrics, rated by two board-certified nuclear medicine physicians, each with over 10 years of clinical experience. A stratified random sample of 40 patients was selected from each of the 10 test groups corresponding to different acquisition durations (3 s to 600 s). Each image was rated on a 5-point Likert scale across four predefined visual quality categories: noise suppression, artifact reduction, contrast retention, and lesion discrimination. Noise suppression evaluates the level of background graininess. Artifact reduction refers to the decrease in typical NAC-related image artifacts, such as edge blurring, halo effects, or focal distortions near high-uptake zones. Contrast retention assesses the preservation of lesion-to-background contrast. Lesion discrimination measures the clarity of lesion boundaries and their visual separability from surrounding tissue. As shown in Fig. 6, the algorithmically enhanced images (blue polygons) consistently outperform the original 2D coronal PET MIP inputs (red polygons) across all time points. The improvement is most pronounced in short-duration scans (e.g., 3 s and 15 s), particularly in noise suppression and artifact reduction, confirming that our deep learning model effectively improves clinical interpretability under ultra-low acquisition settings.Fig. 6A comparison of image quality scores for input and output images across several parameters: noise suppression, artifact reduction, contrast retention, and lesion discrimination. The blue box represents the generated image, while the red box represents the input image. Each aspect is rated on a scale up to 5, indicating the highest possible subjective score A comparison of image quality scores for input and output images across several parameters: noise suppression, artifact reduction, contrast retention, and lesion discrimination. The blue box represents the generated image, while the red box represents the input image. Each aspect is rated on a scale up to 5, indicating the highest possible subjective score Lesion detection performance Figure 7a reports the ROC curves of the four competing methods. The proposed method consistently lies above the UNet, GAN, and DDPM baselines across almost the entire range of false-positive rates, indicating superior sensitivity at matched specificity. Figure 7b further presents the ALROC curves, which incorporate both detection and localization performance. Again, the proposed method achieves the highest curve, demonstrating more accurate lesion localization than the competing approaches. Together, these results confirm that our framework provides the most reliable lesion detectability and localization among all evaluated methods.Fig. 7Lesion-level ROC and ALROC (with localization) curves for UNet, GAN, DDPM, and the proposed method. a Standard ROC curves (case-level lesion detection). b ALROC curves Lesion-level ROC and ALROC (with localization) curves for UNet, GAN, DDPM, and the proposed method. a Standard ROC curves (case-level lesion detection). b ALROC curves Discussion In this study, we developed and evaluated SnapPET, a CT-sparing, deep learning–based framework for attenuation correction and denoising of total-body 2D coronal PET MIP images in pediatric patients. By mapping ultra-short 3-s 2D coronal PET MIP acquisitions to AC-like 2D coronal whole-body MIP images using a latent-diffusion–based autoencoder, SnapPET substantially improves image quality while preserving lesion uptake, and enables rapid on-console reconstruction suitable for clinical workflows. This approach is particularly attractive for pediatric oncology, where repeated PET/CT examinations and the associated cumulative radiation dose—especially from the CT component—are major concerns. Quantitative image-quality metrics demonstrated consistent improvements of SnapPET over the original NAC inputs across all acquisition durations. For ultra-short scans (e.g., 3 s and 6 s), PSNR and SSIM increased markedly, while NMSE decreased, indicating that the model is able to recover substantial structural detail from highly noisy data. ROI-based analyses further showed that SUVmax and SUVmean in lesions reconstructed by SnapPET exhibit strong agreement with the 600-s AC PET reference, with minimal bias. These findings suggest that SnapPET not only improves visual image quality but also maintains quantitative fidelity, supporting its potential role in semi-quantitative assessment and treatment response evaluation. Reader studies provided complementary evidence of clinical utility. The generated images received higher scores than the NAC inputs for noise suppression, artifact reduction, contrast retention, and lesion discrimination, with the greatest gains observed at very short acquisition times. Nonetheless, for the most challenging ultra-low-count cases, subjective scores did not fully reach the level of longer-duration AC PET, indicating that there remains room for improvement in preserving subtle lesion conspicuity and fine anatomical details. The lesion-level ROC and ALROC analyses showed that SnapPET outperformed UNet, GAN, and DDPM baselines in both detection and localization performance, confirming that the proposed latent-diffusion framework offers superior lesion detectability compared with more conventional deep-learning architectures. This advantage is consistent with the design rationale of the revised architecture. In ultra-short-acquisition NAC PET MIPs, severe noise and intensity non-uniformity make direct one-step restoration prone to oversmoothing and loss of subtle lesions. By first learning a compact latent representation and then progressively refining it through conditional diffusion, SnapPET achieves a better balance between noise suppression, lesion conspicuity, and preservation of the overall uptake distribution. This interpretation is further supported by the direct comparison with the previous model provided in the Supplementary Material. From a clinical workflow perspective, SnapPET supports a personalized CT strategy that balances diagnostic information and radiation exposure. A further point worth emphasizing is why we chose AC-like image generation, rather than directly building a lesion-prediction model in the NAC domain, for the triage task. We agree that triage is primarily detection-oriented rather than strictly quantitative. However, under ultra-short acquisitions, whole-body NAC PET MIPs are affected not only by severe count-limited noise but also by attenuation-related intensity non-uniformity, both of which can reduce lesion-to-background conspicuity and increase inter-reader variability. In this setting, generating an AC-like 2D PET MIP provides a more standardized and clinically familiar representation for rapid physician interpretation. A NAC-domain lesion predictor is a reasonable alternative, but it would require an additional task-specific classifier, separate calibration, and a narrower evaluation framework. By contrast, AC-like image generation is task-agnostic, can be directly compared with the 600-s AC reference, and can support both visual reading and downstream analyses. For this reason, we consider AC-like MIP generation to be better aligned with our goal of pre-CT triage and workflow optimization. For pediatric lymphoma patients with negative SnapPET findings, a low-dose CT protocol can be used for basic anatomical localization and attenuation correction, potentially avoiding unnecessary diagnostic CT. For patients with positive SnapPET findings, regional diagnostic CT focused on 18F-FDG-avid regions, combined with low-dose CT elsewhere, can be adopted within a single continuous CT acquisition using bed-specific parameters. In this way, SnapPET is not intended to replace diagnostic CT when detailed anatomical assessment is required, but rather to serve as a triage tool that helps reserve higher-dose CT for cases and regions where it is most clinically justified, in line with the ALARA principle. Several limitations of this study should be acknowledged. First, the dataset was derived from a single center and focused on pediatric 18F-FDG-avid lymphoma; the generalizability of SnapPET to other tumor types, tracers, and adult populations remains to be established. Second, the model operates on 2D MIP representations, which trade detailed 3D anatomical information for speed and simplicity; extending the framework to full 3D volumes or incorporating multi-view projections may further enhance lesion characterization. Third, although 200 patients represent a relatively large pediatric cohort for a diffusion-based model, larger and more diverse multi-center datasets will likely improve robustness, particularly for rare patterns and atypical presentations. Finally, our reference standard was 600-s AC PET rather than clinical outcomes; future prospective studies should evaluate the impact of SnapPET on real-world diagnostic and management decisions. Despite these limitations, our findings indicate that SnapPET can substantially improve image quality and lesion detectability for ultra-short 2D coronal PET MIP acquisitions while preserving quantitative accuracy, outperforming several representative deep-learning baselines. With further validation in larger and more diverse cohorts, and potential extension to other clinical scenarios, SnapPET may help reduce reliance on routine attenuation-correction CT and facilitate personalized, lower-dose CT strategies in pediatric PET/CT, ultimately contributing to safer and more efficient imaging pathways for young patients. Conclusion In this study, we developed SnapPET, a deep learning–based CT-sparing pre-CT triage method that denoises ultra-short 2D NAC PET MIPs to support rapid screening. Rather than eliminating CT, SnapPET guides whether subsequent CT should be low-dose or diagnostic and enables region-adapted CT protocols, which may reduce unnecessary CT-related radiation exposure in pediatric patients. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1