Whole‑body organ segmentation is critical for radiotherapy planning and systemic disease assessment, but existing PET/CT‑tailored segmentation methods suffer from insufficient anatomical coverage, severe annotation bottlenecks, and an unfavorable trade‑off between accuracy and speed, limiting their clinical translation. To address these issues, we developed a weakly supervised framework for whole‑body segmentation of 147 anatomical structures on PET/CT and established normative SUV reference ranges in healthy individuals. This retrospective study included 2308 18F‑FDG PET/CT scans from two hospitals and 1746 public CT scans. Model development used 1846 scans for pseudo‑label training (1746 public CT + 100 PET/CT) and 2000 unlabeled PET/CT scans for weakly supervised learning; independent testing used 110 PET/CT scans. A healthy cohort of 98 individuals was included to establish normal SUV reference ranges. An uncertainty‑guided pseudo‑label learning algorithm was trained to segment 147 anatomical structures. Dice scores were calculated between model predictions and expert annotations. Paired t‑tests or Wilcoxon tests with Bonferroni correction were used for statistical analysis. The model achieved a mean Dice score of 0.874 across 147 structures on the test set, with segmentation completed within 120 s per scan. For 113 structures matched to TotalSegmentator, it outperformed TotalSegmentator (0.895 vs. 0.884; P < 0.001). The uncertainty-weighting strategy improved segmentation of small vascular structures (0.702 vs. 0.651; P < 0.001). The model also showed a lower absolute deviation in the 95th-percentile SUV (SUV95; 0.018 vs. 0.021) than TotalSegmentator in PET quantification. Normal SUV reference ranges for 147 structures were established from the healthy cohort. The proposed weakly supervised framework enables precise and efficient whole-body segmentation of 147 structures on PET/CT. The established normal SUV reference ranges provide a standardized benchmark for quantitative PET/CT interpretation. The source code and pretrained weights are publicly available at https://github.com/Wpx01/WiseSeg.
Attenuation in PET imaging can lead to quantification errors, underestimation of lesions and misinterpretation of the extent of disease. Commonly, this correction is made using a CT image that provides additional anatomical information. Unfortunately, using CT significantly increases the radiation dose received by patients, increases the cost of imaging equipment, and can cause misregistration errors. To overcome these limitations, deep-learning based methods were proposed to generate a pseudo-CT from the non-attenuation-corrected (NAC) PET image, which can then be used to correct for attenuation. In this work, an improved UNet 2.5D architecture has been developed in which three consecutive PET image slices are used to provide more contextual information to generate a single CT slice. A dataset of 189 NAC head PET scans from patients injected with ^18 F-fluorodeoxyglucose or ^11 C-acetoacetate, and the corresponding CT images were used for training. The generated pseudo-CT of all patients showed an overall improvement with the proposed UNet 2.5D architecture. Compared to a standard UNet 2D architecture, the soft tissue dice score increased from (0.905 ± 0.015) to (0.922 ± 0.017), along with the SSIM score increasing from (0.889 ± 0.023) to (0.921 ± 0.019). 3D forward projections comparing the CT and pseudo-CT total attenuation projection values also highlight the improved attenuation maps generated by the proposed 2.5D deep learning architecture. The proposed 2.5D UNet, using adjacent PET image slices for added context, outperforms the 2D UNet in generating accurate pseudo-CTs from NAC PET images, enabling effective and fast attenuation correction without additional radiation or imaging.
Robust single-time-point (STP) dosimetry critically depends on selecting an imaging time that yields reliable estimates of the time-integrated activity coefficient (TIAC). This study developed and evaluated a virtual-patient nonlinear mixed-effects modelling (NLMEM) framework to optimise STP imaging schedules for renal TIAC estimation, using [¹⁷⁷Lu]Lu-PSMA-617 therapy as a proof of concept. A previously published NLMEM (sum-of-exponentials structure and parameter distributions) describing renal [¹⁷⁷Lu]Lu-PSMA-617 biokinetics in 63 patients served as the generative model [1]. Based on fixed- and random-effects parameters, 500 virtual patients (VPs) were sampled, and reference TIAC values were computed analytically (rTIAC). For each candidate imaging time, renal activity measurements were simulated by applying proportional noise (7.9
To develop an anthropomorphic phantom that enables the simultaneous evaluation of key performance metrics for nuclear medicine imaging systems and to preliminarily validate its application in PET/CT. The anthropomorphic phantom replicates thoracic and abdominal anatomy, simulating human skeletal structures, lung tissue, soft tissue, and tumors. The phantom is divided into five functional regions: a lung tumor detection region (right lung with spheres), an attenuation and scatter correction assessment region with a non-radioactive compartment (left lung), an abdominal tumor detection region (right abdomen with spheres), a uniform background region (left abdomen), and a spatial resolution assessment region incorporating line sources, centered along the mid-sagittal plane traversing the thoracoabdominal extent. All phantom imaging was performed on a Discovery MI (GE Healthcare) PET/CT scanner using sphere-to-background ratios (SBRs) of approximately 4, 6, 8, and 10, and line-to-background ratios (LBRs) of approximately 200, 500, 1000, and 2000. Spatial resolution was measured using a two-point-source method, including visual assessment and valley-to-peak ratio (VPR) analysis, and a single-point-source method. Lesion detectability was assessed visually and by contrast-to-noise ratio (CNR). Signal-to-noise ratio (SNR), quantitative accuracy (percentage error, PE), image uniformity (IU), and residual error (RE) for attenuation and scatter correction were measured. Additionally, a matched SBR comparison was performed using the NEMA image quality (IQ) phantom. As LBR increased from 200 to 2000, the single-point-source spatial resolution improved from approximately 7 mm to 3 mm, whereas the visually determined two-point-source results remained nearly constant around 6 mm. At the fixed 7-mm spacing, VPR decreased with increasing LBR, indicating improved profile separation between adjacent point sources. The 3 mm sphere was detectable in the lung and abdomen at SBRs of ≥ 6 and 10, respectively. For SBRs set to approximately 4, 6, 8, and 10, SNRs measured in the left abdominal uniform region were 20.50, 25.47, 25.19 and 25.75; PE, axial IU, and RE were all < 5
This study evaluated whether delayed static 18F-flutemetamol PET acquired at 60 min post-injection on digital SiPM PET/CT can be quantitatively bridged to the conventional 90-min Centiloid (CL) scale while preserving visual interpretability. Methods: In the derivation cohort, 63 participants underwent paired 60- and 90-min static 18F-flutemetamol PET acquisitions on digital SiPM PET/CT systems. Three experienced readers independently performed visual assessment (VA) at each time point. Raw quantitative agreement between acquisition windows was first evaluated using SUVR and direct VIZCalc-derived CL values. A study-derived SUVR₆₀-to-CL₉₀ conversion equation was then developed and evaluated using Bland–Altman analysis, Lin’s concordance correlation coefficient (CCC), and leave-one-out cross-validation. The same equation was additionally applied without refitting to an independent internal validation cohort of 34 paired examinations. Results: The majority VA was concordant between 60 and 90 min in 62/63 examinations (98.4
I-131 radioactive iodine (RAI) is widely used in thyroid cancer patients with bone metastases (BM). While blood-based dosimetry has been standard for predicting marrow toxicity, tumour dosimetry remains an unmet need. We evaluated the feasibility of pretherapeutic quantitative I-131 SPECT/CT for patient-tailored dosimetry in poorly differentiated thyroid carcinoma (PDTC) patients with BM. Five PDTC patients (18 BM lesions) underwent serial pre-therapeutic I-131 SPECT/CT (111 MBq, eight time points up to 120 h) with concurrent blood sampling. I-131 concentrations (MBq/g) were quantified using quantitative SPECT/CT technology and validated against gamma counting. Absorbed dose rates (Gy/h) for blood and tumours were estimated using Monte Carlo simulations (customized Geant4, cross-validated by TOPAS). Tumour absorbed doses were calculated via trapezoidal integration, and clinical outcomes were assessed after subsequent high-dose RAI therapy. Blood I-131 concentrations measured by SPECT/CT showed excellent agreement with gamma counting (p < 0.001). Blood absorbed doses per administered activity were comparable between SPECT/CT and conventional methods (p = 0.6250). Tumour I-131 concentration strongly correlated with absorbed dose rate, resulting in a regression equation [tumour absorbed dose rate (Gy/h) = 0.1256 × tumour RAI concentration (MBq/g) (R2 = 0.9918, p < 0.0001)]. Reducing imaging time points from 8 to 4 maintained the accuracy of absorbed dose estimation. Although exploratory, tumours with expected absorbed doses > 50 Gy after high-dose RAI showed no progression, whereas lower-dose lesions progressed more frequently (87.5
Interictal SPECT is a critical component of SISCOM analysis for presurgical localisation of epileptogenic regions. However, its clinical reliability can be affected by uncertainty regarding whether the interictal scan is a true representation of a seizure free baseline (i.e., the interictal state). In addition, most patients will also have an interictal FDG PET scan and MRI performed, and it would be useful if the number of scans required could be reduced. This study explores the use of machine learning synthesised interictal SPECT within the SISCOM framework (i.e., SySCOM) as an alternative baseline to the acquired interictal SPECT scan. PET scans from 40 patients with refractory focal epilepsy were used to train a three-dimensional generative adversarial network to synthesise interictal SPECT images. Synthesis performance was evaluated on an independent validation cohort of nine unseen patients using standard image quality metrics (i.e., SSIM, PSNR, RMSE). Both real and synthetic interictal SPECT images were used to generate SISCOM outputs following standard clinical protocols. Five clinicians experienced in the interpretation of SPECT images in epilepsy performed blinded visual localisation of regions of hyperperfusion across 18 predefined anatomical regions. Observer based similarity, agreement, and reproducibility between SISCOM and SySCOM were assessed. Synthetic interictal SPECT images demonstrated high similarity to real images (mean SSIM = 0.94, mean PSNR = 27.9 dB, mean RMSE = 0.041). Visual localisation based comparison of SISCOM with SySCOM demonstrated good agreement between the methods (Dice score = 70
Although multiple PET reconstruction approaches have been successfully applied in brain and oncologic imaging, their performance in the context of spinal cord imaging remains unverified. This study aims to identify the optimal reconstruction method for spinal cord PET image analysis based on assessments of image quality and quantitative uptake. We retrospectively analyzed 18F-fluorodeoxyglucose (FDG) PET/MR data from 22 patients with cervical spinal cord tumors. PET images were reconstructed using Ordered Subset Expectation Maximization (OSEM), OSEM + Point Spread Function (PSF), OSEM + Time-of-Flight (TOF), and OSEM + TOF + PSF algorithms. Image quality was assessed by noise, signal-to-noise ratio (SNR), and signal-to-background ratio (SBR), both overall and within equally sized small (n = 11) and large (n = 11) tumor subgroups. Tumor uptake metrics, including SUVmean, SUVmax, SUVRmean, and SUVRmax of the lesions, were quantified for each method. The differences among reconstruction methods were assessed using the Friedman test. There were no significant differences in image noise among reconstruction methods (P > 0.05). Compared to OSEM reconstruction, the OSEM + PSF + TOF method significantly improved SNR (Z = 3.62, P < 0.01) and SBR (Z = 2.80, P = 0.03); the SBR of OSEM + PSF + TOF was also significantly higher than that of OSEM + PSF (Z = 3.04, P = 0.01). In large tumors, while image noise remained unaffected by the reconstruction method (P > 0.05), both SNR and SBR were significantly impacted (both P < 0.05). Specifically, OSEM + TOF + PSF achieved significantly higher SNR compared to OSEM (Z = 3.30, P < 0.01), and higher SBR compared to OSEM + PSF (Z = 3.14, P = 0.01). Quantitative lesion uptake metrics (SUVmax, SUVmean, SUVRmax, and SUVRmean) were significantly higher for OSEM + TOF + PSF compared to both OSEM and OSEM + PSF reconstructions (all P < 0.05). Additionally, SUVmean (Z = 2.92, P = 0.02) and SUVRmean (Z = 2.83, P = 0.03) values from OSEM + TOF + PSF reconstruction were significantly higher than those from OSEM + TOF. The OSEM + PSF + TOF reconstruction method demonstrated significant advantages in both image quality and quantitative assessment of FDG uptake, making it the recommended choice for spinal cord PET image reconstruction.
Abstract Background Radionuclide therapy is an increasingly used treatment modality for patients with disseminated cancer diseases. Tumor-binding radiopharmaceuticals labelled with the alpha-particle emitter astatine-211 ( 211 At) are of interest to enable high radiation absorbed doses to small tumors. The alpha particles emitted by 211 At will deposit all their energy in a short-range including neighboring cells, causing localized irradiation. Due to being part of group 17 the halogen family with iodine, free 211 At accumulates in the thyroid gland, which is the main normal organ at risk. It is therefore of great importance to estimate the radiation absorbed dose accurately to the thyroid gland for free 211 At. The aims of the work were to determine microdosimetric quantities for uniform and non-uniform 211At distributions within different models of the thyroid gland using the Monte Carlo (MC) codes AlfaMC and GEometry ANd Tracking (Geant4), and to compare the results with previously published data using Monte Carlo N-Particle eXtended (MCNPX). AlfaMC is customized for calculating microdosimetric quantities for alpha-particles. Results Both a single and a multiple thyroid follicle model designed for man, rat and mouse were used from previously published studies. Monte Carlo simulations and absorbed dose-calculations were performed for different distributions of 211At within the thyroid models. The mean specific energy, single-hit mean specific energy and single-hit specific energy distribution were calculated for follicle cell nuclei in models for each species. Conclusion The overall agreement between AlfaMC and MCNPX results, and between AlfaMC and Geant4 results, was found to be within 5% and 2%, respectively. The intercomparison suggested an effective way for validation of microdosimetry results; of AlfaMC validation, a code designed with a focus on reducing the calculation time.
Transarterial radioembolization (TARE) is evolving from palliative to curative in the management of hepatocellular carcinoma (HCC). Alongside, there is a growing need to develop personalized and predictive dosimetry strategies to maximize the therapeutic effect. The objective of this study was to evaluate the feasibility of a novel framework, MIDOS, to design patient-specific optimal TARE plans. A retrospective cohort of fourteen TARE patients with solitary HCC who had [99mTc]Tc-MAA SPECT-CT was analyzed, including segmental and lobar administrations. For each patient, tumor and perfused area segmentations were included in a liver vasculature model. In our simulations, microspheres were stochastically assigned in cluster patterns to tumor or normal liver tissue, with probabilities proportional to flow and modulated by tumor and normal liver uptake models. Simulation results were employed to generate synthetic [99mTc]Tc-MAA SPECT images that were compared with patient images using the relative activity errors in tumor and normal tissue, tumor-to-normal ratio (TNR), and gamma analysis (10
Abstract Purpose To date, some studies have employed deep learning techniques to directly generate dynamic positron emission tomography (PET) parametric images from static PET. Compared with traditional methods, this approach requires only a single PET/computed tomography (CT) scan. However, current methods tend to employ static PET images captured at fixed scanning times without considering whether static PET images acquired at different scanning times impact the quality of the dynamic PET parametric images generated via deep learning. Methods A single-frame image from dynamic $$^{68}{\textrm{Ga}}$$ -FAPI-04 total-body PET can actually be regarded as a static PET image acquired during that specific period of the PET/CT scan. We extracted 5 frames of dynamic $$^{68}{\textrm{Ga}}$$ -FAPI-04 total-body PET at equal intervals, specifically frames 50, 76, 80, 86, and 92. Each frame of the dynamic $$^{68}{\textrm{Ga}}$$ -FAPI-04 total-body PET images was subsequently input into the deep learning model to obtain dynamic $$^{68}{\textrm{Ga}}$$ -FAPI-04 total-body PET parametric images. By comparing and analyzing the quality of dynamic PET parametric images, we attempted to determine the optimal scan time for static PET. Results The experimental results revealed that the dynamic $$^{68}{\textrm{Ga}}$$ -FAPI-04 total-body PET parametric image generated from the 58th frame of the dynamic PET image via the deep learning model exhibited the poorest performance. This may be because the diffusion of the radioactive tracer was not stable at the time of PET/CT scanning in frame 58, but it was relatively stable from frames 70–92. Conclusion Static PET images acquired at different scanning times do indeed affect the quality of the dynamic $$^{68}{\textrm{Ga}}$$ -FAPI-04 total-body PET parametric images generated via deep learning. Specifically, when the radioactive tracer is unstable during the early stage of scanning, the quality of the dynamic $$^{68}{\textrm{Ga}}$$ -FAPI-04 total-body PET parametric images generated from static PET images via deep learning appears to be inferior.
The comprehensive knowledge of how the distinct physical decay characteristics of each radionuclide shape their image quality performance is essential for tailoring acquisition protocols and ensuring optimal clinical applications. Understanding the spatial dependence of the contrast recovery coefficient (CRC) is a prerequisite for robust quantitative image analyses and accurate clinical interpretations, since the partial volume effect (PVE) and the count rate vary along the axial field of view (AFOV). Image quality assessments with various radionuclides were performed on the uMI Panorama GS PET/CT system, with the CRC profile across its extended AFOV also evaluated. Following the NEMA NU 2-2018 standard, the image quality was assessed using the NEMA IEC (NEC) Body Phantom with 18F, 64Cu, 68Ga, 89Zr, 124I, and 90Y. The CRC, the background variability (BV), and the lung residual error were measured to quantify the image quality. CRCs were sequentially measured at four axial offsets: 1/2, 1/4, 1/8, and 1/16 of the AFOV. The impact of the radionuclide-specific PSF (point spread function) modeling on the CRC was evaluated by comparing PET images reconstructed with and without PSF modeling. The assessment demonstrated comparable image qualities for 18F, 64Cu, 68Ga, 89Zr, 124I, while 90Y showed a similar CRC but a higher BV and a higher lung residual error, attributable to its low positron branching ratio. The BV and lung residual error increased with increasing axial distance from the AFOV center, whereas CRCs for larger spheres remained stable across AFOV positions; small-sphere CRC measurements exhibited increased dispersion toward the periphery due to reduced count statistics. The PSF modeling significantly improved CRCs for 68Ga and 124I, consistent with their long positron ranges. All investigated radionuclides were effectively imaged on the uMI Panorama GS PET/CT system. The AFOV uniformity was quantitatively evaluated based on CRC measurements obtained at multiple positions along the AFOV.
PET images often make small lesions difficult to identify because of noise and system blur. We address this by developing and evaluating MlPET , a fast localized machine-learning method that approximates a computationally expensive probabilistic sampling approach while reducing noise and increasing spatial resolution. Building on a probabilistic deconvolution framework with informed priors, MlPET replaces computationally demanding Markov chain Monte Carlo sampling with a localized neural network trained to directly estimate the posterior mean of voxel activity from small image neighborhoods. The method incorporates scanner-specific point spread functions (PSF), spatially correlated noise modeling, and flexible prior information. Performance was evaluated on NEMA phantom data acquired on three PET systems (GE Discovery MI, Siemens Biograph Vision 600, and Siemens Biograph Vision Quadra Edge) under varying reconstruction settings and acquisition times. On NEMA IEC phantom data, MlPET produced contrast-recovery coefficients consistently higher than PET and frequently close to 1.0 (including the 10 mm sphere in several settings), while simultaneously reducing background noise and improving spatial definition. The effective point-spread function (PSF) full width at half maximum (FWHM) was on average reduced from about 2 mm in PET to below 1 mm with MlPET , corresponding to roughly a 2.5× decrease in effective blur. Comparable image quality was obtained at 40–80 s acquisition time using MlPET versus 900 s with conventional PET. A clinical example in a breast cancer patient illustrates potential clinical applicability. MlPET provides a computationally efficient approach for quantitative probabilistic post-reconstruction PET image analysis. By combining informed priors with the speed of a neural network, it achieves both noise suppression and resolution enhancement. The method shows promise for improved small-lesion detectability and quantitative reliability in clinical PET imaging. Future clinical studies will evaluate its performance on patient data and quantify effects under realistic prior uncertainty.
Positron emission tomography with magnetic resonance imaging (PET/MRI) provides noninvasive molecular characterization of breast cancer and has the potential to improve diagnostic accuracy, staging, treatment response assessment, and guide personalized care. Reducing the radiation dose from 2-deoxy-2-[18F]fluoro-D-glucose (18F-FDG) to a level similar to digital mammography while maintaining image quality may facilitate clinical utilization. This study was performed to evaluate diagnostic image quality and lesion conspicuity of low-dose 18F-FDG breast PET/MRI using denoising and to evaluate the effect of denoising on radiotracer uptake semi-quantification. This pilot study was a secondary analysis of a single-institution prospective study of 18F-FDG breast PET/MRI for 23 women with primary invasive breast cancer. Random undersampling of the PET data from a 30-min simultaneous prone 18F-FDG breast PET/MRI was used to produce simulated low-dose (SLD) images approximating 90
To evaluate the radiation protection requirements for a UK first specimen PET-CT scanner, the AURA10 (Xeos, Ghent), as assessed through a clinical trial investigating its use for intra-operative surgical margin assessment in urology and head and neck (H N) tumour removal. A radiation risk assessment was performed to inform radiation protection requirements and training. 0.8 MBq/kg of 18Fluorine fluorodeoxyglucose and 50 MBq of 68Gallium prostate-specific membrane antigen were administered for H N and urology cases respectively. Whole body optically stimulated luminescence dosimeters, finger and eye ThermoLuminescent dosimeters and electronic personal dosimeters were issued to staff in theatres, recovery, and histopathology. Dose rate measurements were recorded at the position of staff members in theatres. Measured counts and dose rate from the resected specimen and swabs were obtained and contamination monitoring of the room were performed following removal of the patient. Training sessions and dry runs with surgical staff helped identify radiation risks and optimise workflow. Occupational doses from surgeons and scrub nurses were the highest for both modalities. Doses to other theatre staff were below dose limits for members of the public in the UK. Doses to staff outside theatres, including recovery and histopathology, were found to be below threshold values. Low level contamination was present in swabs (< 6 kBq and < 250 kBq for H N and urology cases respectively). Thorough staff training, systems of work and shielding are required to ensure radiation doses to staff are minimised and contamination and radioactive waste are contained. There were difficult occasions during shift changes and untrained staff rota-ed in, but nuclear medicine supervision and then the introduction of theatre-based cascade trainers helped solidify learning and provide assurance to theatre staff and staff showed good compliance with the systems of work in place. Annual doses to theatre staff are heavily dependent on workload, with surgeons and scrub nurses closest to the patient receiving whole body doses of 39 µSv and 12 µSv, and finger doses of 88 µSv and 48 µSv respectively per case for H N. Corresponding urology doses per case were measured as 7.2 µSv and 2.5 µSv for whole body and 16 µSv and below threshold values for the assisting surgeon and scrub nurse respectively. www.clinicaltrials.gov. URL trail registry record: Study Details | NCT06676943 | Investigating the Diagnostic Performance of High-resolution Specimen PET-CT in Determining Margin Status in Cancer Resection | ClinicalTrials.gov.
To develop a data-efficient, unsupervised deep learning framework for deformable SPECT/CT registration that supports voxel-based dosimetry in radionuclide therapy, particularly in clinical settings with limited imaging datasets. We propose Nufit-Reg, an unsupervised deep learning network for deformable SPECT/CT registration. The network employs dual Swin-Transformer-based encoders to process CT and SPECT images separately, with cross-stitch units enabling structured feature sharing and multimodal integration across hierarchical stages. To overcome data scarcity, a two-stage training strategy was adopted. In the first stage, inter-patient pre-training was performed using randomly paired iodine-131 SPECT/CT scans (a total of 58 scans) to learn general anatomical correspondences. In the second stage, few-shot fine-tuning was performed using intra-patient sequential SPECT/CT pairs derived from quantitative scans (12 patients, each with three time points), enabling the model to adapt to patient-specific temporal consistency while being optimized across multiple patients. Registration performance was evaluated against Elastix and UTSRMorph using image similarity metrics, including structural similarity (SSIM), weighted SSIM (wSSIM), local normalized cross-correlation (LNCC), and mutual information (MI), as well as deformation regularity assessed via Jacobian determinant analysis. The downstream impact of registration on voxel-level time-activity-curve (TAC) fitting and tumor dose distributions (D2
¹⁸F-MK6240 tau positron emission tomography (PET) is critical in Alzheimer’s disease diagnosis and therapeutic monitoring. However, the standard 20-minute static acquisition poses challenges for cognitively impaired patients with limited tolerance for prolonged scans. This study evaluates the feasibility of reducing scan time while maintaining diagnostic accuracy through image quality and quantitative comparisons. 69 memory clinic patients (27 males and 42 females) who underwent 18F-MK6240 PET examinations were retrospectively analyzed. All 20-minutes scans were acquired 90 min post-injection of 3.7 MBq/kg ¹⁸F-MK6240 and reconstructed mainly into 5-, 10-, 15-, and 20-minute datasets. Images were rated on a 5-point scale for overall quality, noise, and diagnostic confidence. Tau status (positive/negative) was the visually defined, and standardized uptake value ratios (SUVr) of Braak regions were calculated using cerebellar gray matter as reference. Agreement across durations was assessed via Bland–Altman analysis. Among the 69 patients, 50 were classified as tau-positive and 19 as tau-negative. Diagnostic efficacy remained consistent across acquisition times compared to the 20-minute reference. The slight reduction in image quality for the 5-minute images relative to the other groups, but none of the images across any of the four durations were rated as poor (score < 3). For 16 predefined regions used for visual assessment of regional tau involvement, both readers showed near-perfect agreement in scoring the extent of involvement (0
The aim of this study was to optimise the acquisition and reconstruction parameters for ^90Y imaging on a digital PET scanner (Discovery MI 4-ring) under high (3 GBq), intermediate (1 GBq), and low (200 MBq) activity conditions. First, quantitative linearity was evaluated. Then, NEMA IEC body phantom acquisitions were reconstructed using various scan durations and Q.Clear reconstruction parameters. Next, optimal protocols were identified by minimising the discrepancies between absorbed dose maps derived from images (using the Local Deposition Model (LDM)) and those obtained from Monte Carlo (MC) simulations. The images reconstructed with these optimal protocols were then used to evaluate effective spatial resolution and to compare the accuracy of LDM with that of the Dose Voxel Kernel (DVK) approach. Quantitative linearity analysis revealed an underestimation of phantom activity at high activity (up to − 15.7
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. 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. 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). 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.
Abstract Purpose The aim of this study was to investigate the limits of diagnostics and therapy planning for patients with prostate cancer using non-time-of-flight 18F/68Ga-PSMA PET/MRI under clinically challenging imaging conditions with small lesion sizes and low uptake. Lesion detectability and quantification accuracy were evaluated for different acquisition and reconstruction parameters in a systematic phantom study and subsequent on patient data. Methods PET/MRI measurements were performed using a small lesion NEMA phantom. PET data were acquired for nine different activity concentrations (AC). Data of a longer single-bed protocol in the pelvis or a shorter whole-body protocol were reconstructed using relative or absolute scatter correction (SC). PET images were analysed considering a ± 25% deviation range between imaged and true AC as acceptable. Thirteen PSMA-PET/MRI patients with primary lesions or lymph node metastasis < 12 mm in the pelvis were included in this study. The presence of the halo artefact was evaluated in six 18F-PSMA and seven 68Ga-PSMA PET/MRI patients. For 21 lesions (diameter 6.4–12.3 mm) in total, the AC was quantified. Results For both radiotracers, the 9.7 mm sphere was still visible at 0.16 kBq/mL with emission times > 40 min. The 3.7 mm sphere was only detectable at 22 kBq/mL with emission times > 4 min. All spheres ≥ 6.5 mm provide acceptable quantification at an AC of 1.32 kBq/mL for 18F PET/MRI protocols of ≥ 12 min and 2.75 kBq/mL for 68Ga PET/MRI protocols. In phantom data, no halo artefact was observable and different SC methods had no impact on quantification. 4/6 18F-PSMA patients and 7/7 68Ga-PSMA patients showed a halo around the bladder using relative SC, which could be reduced in all patients using absolute SC. Comparing the minimum quantifiable AC (MQAC) from in the phantom study as a threshold to the patient data, all lesions provided acceptable quantification with values > MQAC (AC 3.3–108.5 kBq/mL) for equal reconstruction and acquisition parameters. Conclusions The results demonstrated that the detection of lesions in the sub-centimetre range and a reliable quantification of 18F/68Ga-PSMA uptake using standard acquisition and reconstruction parameters within clinical PET/MRI protocols is possible. This allows for an individual assessment of potential therapy options for each patient.
