Introduction Positron Emission Tomography (PET) image reconstruction is traditionally performed using analytical or iterative algorithms, which can be time-consuming and prone to increased noise in low-count frames. Deep learning methods, such as FastPET, offer rapid reconstruction with improved noise suppression. This study aimed to evaluate the quantitative performance of a revised FastPET approach for dynamic whole-body 82Rb PET imaging on a long axial field-of-view (LAFOV) scanner, using iterative PSF-TOF reconstruction as the reference standard. Methods Forty-one patients underwent rest–stress 82Rb PET/CT scans on a LAFOV PET/CT system. List-mode data were reconstructed into 35 dynamic frames and static images corresponding to last 3 min using both FastPET (3D U-Net with residual blocks) and PSF-TOF OSEM. Image quality was assessed using signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR). Organ uptake values, time activity curves (TACs), and the tracer blood to tissue influx rate (i.e. K1) from a one-tissue compartment model were compared between methods. Results Static FastPET images showed 2.4 times higher SNR and 2.3 times higher CNR at rest, and it was 2.8-times higher for both metrics under stress conditions (p < 0.001). Organ uptake values measured from FastPET images correlated strongly with values derived from PSF-TOF images (R2 = 0.823–0.997). Absolute TAC area under the curve differences were less than 6% for all organs and conditions. No significant differences were observed in regional K1 values between methods for myocardium, kidneys, spleen, or other regions. Conclusion FastPET enables accurate, high-quality dynamic 82Rb PET reconstruction with approximately a 70% reduction in computation time compared to standard PSF-TOF reconstruction. It preserves quantitative parameters and temporal dynamics while providing enhanced SNR and CNR, supporting its potential application in dynamic 82Rb PET imaging workflows.

The image quality and quantitative accuracy of 82Rb myocardial perfusion imaging (MPI) using PET is challenged by the extensive positron range (PR) effects, with the PR of 82Rb being about 7 mm in soft tissues. This study explored the feasibility of applying postacquisition PR correction (PRC) to routine 82Rb PET/CT MPI acquisitions and assessed its impact on diagnostic accuracy and image quality. Methods: We implemented a PRC method adjusted to 82Rb into a vendor-provided reconstruction toolbox, using tissue-specific corrections for soft tissue, bone, and air/lungs. The PRC was evaluated in 2 cohorts: the first comprised 25 healthy volunteers who underwent repeated 82Rb MPI within 2 wk, and the second included 66 patients with known or suspected coronary artery disease. We measured the signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR) for the volunteer cohort. In the patient cohort, the impact of PRC was evaluated as changes in the area under the receiver operating characteristic curve (AUC), using fractional flow reserve as the gold standard (values < 80% were considered significantly reduced). We calculated AUCs for stress and ischemic total perfusion deficits. Results: In the volunteer cohort, PRC-based reconstructions (standard reconstruction [STD] + PRC) demonstrated significantly improved SNR and CNR compared with STD, with median increases of 22% and 47% for SNR and CNR, respectively (P < 0.05). For the patient cohort, comparable AUCs were reported for STD- versus PRC-based reconstructions (stress total perfusion deficits, 0.84 vs. 0.83 [P = 0.49]; ischemic total perfusion deficits, 0.87 vs. 0.87 [P = 0.80]). Conclusion: PRC significantly enhances SNR and CNR compared with STD without affecting the diagnostic accuracy of the scans. Given the significantly improved image quality, PRC may be recommended for MPI using 82Rb PET/CT clinical-routine-assessment interpretation of TPD.

End-to-end deep learning PET reconstruction significantly surpasses traditional iterative methods in speed and shows promise for surpassing them in specific scenarios, such as low-dose imaging. In 2019, a significant advancement was made by using histo-images instead of TOF sinograms as the networks input. Histo-images, by leveraging the images geometry, are more compatible with convolutional neural networks than TOF sinograms. Typically, the networks input comprises a PET data histo-image patch and an attenuation map patch. However, this method has shown inconsistent bias in the reconstructed images. This work demonstrates that bias present in the prior method can be mitigated with alternative representations of attenuation information. Instead of using the attenuation map directly, we propose using a multi-view histo-image of the attenuation correction factors, inspired by the iterative DIRECT framework and standard statistical modeling practices. We tested using them as separate channels, as well as using them to pre-correct the data, or both. This histo-image encompasses the attenuation properties of each voxel from all directions within the entire lines of response. Our approaches significantly enhances image quantification, reducing the relative difference from MLEM to an average of 2.0% to 3.0% across 16 regions of interest, compared to 9.1% with the previous method. Our statistical hypothesis test showed that the proposed methods significantly reduced absolute bias compared to the previous method, with p-values ranging from 0.002 to 0.007.

Deep Learning (DL) PET reconstruction offers a promising alternative to full model iterative reconstructions when computational resources are limited. Previous work introduced the FastPET network, utilizing histo-images instead of projection data as DL network input, allowing for efficient network parameter optimization. However, subsequent investigations indicated that for optimal image representation, normalization and attenuation corrections should be applied to these data, and multi-view histo-images may be necessary. This raises the question of whether highly compressed TOF projection data, used in full model reconstruction, can serve as DL network input. This study explores the feasibility of using histoprojections, compressed TOF data, as DL network input, specifically in the context of Siemens reconstruction software and its highly optimized histo-projections data format. Initial results suggest that while backprojected images from histo-projections may be smoother compared to histo-images, they can be effectively deblurred to a gold standard image using the original FastPET network.

Maximum likelihood expectation-maximization (MLEM) and it is variant ordered-subsets EM (OSEM) have to be terminated in early iterations to reduce noise. Maximum a posteriori (MAP) reconstruction can reach convergence with improved image quality, but it is time consuming. In this study, we trained a conventional convolutional neural network (CNN) as a smart postreconstruction filter to transform the clinical OSEM reconstructed images to converged MAP reconstructed images, using relative difference as penalty function. Quantitative comparisons show that the method can improve lesion contrast as well as reduce background noise.

Respiratory motion correction often encounters prolonged processing times due to iterative reconstruction processes. The emergence of the deep learning framework FastPET presents a promising solution, offering rapid reconstruction while maintaining high quality quantification. In this study, we develop a FastPET-based respiratory motion correction method and validate its efficacy through Monte Carlo simulations with the XCAT phantom. The results show that the FastPET-based motion correction effectively recovers motion blurring in liver and lung tumors. This application and validation of FastPET in motion correction represent significant progress. Future studies will explore various approaches to further optimize the motion estimation and correction techniques.

End-to-end deep learning PET reconstruction has gained popularity in recent years. However, it raises concerns about hallucinations (when the network creates artificial features), generalization capabilities, and the use of imperfect labels. Previous works have proposed using an unsupervised loss function operating on the sinogram domain. However, neural networks are usually trained with image patches to allow faster training and limit the memory requirement. Projecting these patches into sinograms is slow and under-optimal as it involves large sinograms. Replacing them with histo-images enables us to forward project a small image patch into a small histo-image patch very efficiently. Moreover, some studies have shown the superiority of histo-images over sinograms as the data representation input of the end-to-end reconstruction network. In this work, we implemented and evaluated this possibility. We found that adding the data loss in the training process on top of the image loss only slowed the training by a factor of three for the same number of epochs on a long-axial field-of-view scanner, which is very reasonable given the complexity of the forward projection process. Using the pure data loss function gave us a trend similar to the one observed in unpenalized iterative reconstruction, with the noise level increasing simultaneously with the contrast. When using a combination of data and image loss functions, the noise quickly reached a limit over which it was not growing anymore while the contrast was still increasing.

Accurate scatter correction is crucial for quantitative positron emission tomography (PET). While the gold standard, Monte Carlo simulation, is usually too slow to be used routinely, faster alternatives often come at the cost of lower accuracy. To avoid this trade-off between accuracy and computational performance, deep learning-based approaches have recently indicated great potential. Here, we focus on the latest generation of PET scanners and provide an extension of prior work that is dedicated to time-of-flight PET in long axial field-of-view scanners. In particular, we train a U-net-like neural network to reproduce the outcome of our in-house Monte Carlo (MC) simulation as a function of the TOF PET emission data and the CT-based PET attenuation correction factors. Trained on data of 23 patients and tested on another 3 patients, our deep scatter estimation (DSE) yields scatter distributions that differ by 6.4% from the MC ground truth and clearly outperform single-scatter-simulations (error of 21.3%). These errors translate to 6.7% (DSE) and 24.6% (SSS) for scatter-corrected PET reconstructions, demonstrating the potential of DSE for fast and accurate scatter correction in clinical practice.

End-to-end deep learning PET reconstruction is fast and could, in the future, surpass iterative reconstructions in specific cases, such as low-dose imaging. A breakthrough was reached in 2021 using histo-images instead of sinograms as the network input. Histo-images are an alternative format to sinograms that contain the same information but with the geometry of the image, making them much easier to handle by a convolutional neural network. The network input typically consists of a (multi-view) histo-image PET data patch and another attenuation map patch. However, inconsistent bias has been observed in the image and remains unsolved until now. In this work, we show that the leading cause of this bias is the lack of information provided to the network regarding the attenuation of regions outside the patch. This information is indeed necessary for a correct quantification. We propose to replace the attenuation map with a multi-view histo-image of the attenuation correction factors, similar to what is done in the iterative DIRECT framework. Such a histo-image contains the attenuation of each voxel in each direction in the entire lines of response. We propose to use the same angular sampling for the attenuation information and the PET data, as would be done in iterative reconstruction. We achieved a significant improvement in image quantification, with less than 7.7 % relative difference from MLEM in 75 % of slices, compared to 33 % with the previous approach. Our technique also removes some artifacts in the image. The remaining inconsistencies in the quantification could be due to an early stop in the training process and an inaccurate scatter estimation.

LSO background radiation is currently used to monitor scanner detector properties, such as energy spectrum potential drift. However, monitoring time properties remains a challenge because LSO emits photons with lower than 511 keV energies, resulting in different detector responses.An attractive feature is monitoring detector properties through LSO data during patient scans. This can be achieved by acquiring data with two energy windows: one set up for 511 keV emission and another for lower energies of LSO background radiation. In principle, patient data can be self-time aligned, but this is relatively expensive and requires multiple data reconstructions. Moreover, the estimation precision of detector time offsets (TO) can suffer from patient attenuation map misalignment. On the other hand, the LSO background radiation TOF spectrum depends on detector geometry only, and assuming relatively good time alignment can result in a simpler method of TO estimation.Patient data investigations have shown that LSO TOs differ from 511 keV emission TOs. However, due to the simultaneous acquisition of LSO and 511 keV emission data, the LSO footprint can be constantly learned from reliable patient scans, as the true time offsets for 511 keV emission are known from self-time alignment. The learned pattern can be predictive for scanner performance during patient scans.

Attenuation correction is a critically important step in data correction in positron emission tomography (PET) image formation. The current standard method involves conversion of Hounsfield units from a computed tomography (CT) image to construct attenuation maps (µ-maps) at 511 keV. In this work, the increased sensitivity of long axial field-of-view (LAFOV) PET scanners was exploited to develop and evaluate a deep learning (DL) and joint reconstruction-based method to generate µ-maps utilizing background radiation from lutetium-based (LSO) scintillators. Data from 18 subjects were used to train convolutional neural networks to enhance initial µ-maps generated using joint activity and attenuation reconstruction algorithm (MLACF) with transmission data from LSO background radiation acquired before and after the administration of 18F-fluorodeoxyglucose (18F-FDG) (µ-mapMLACF-PRE and µ-mapMLACF-POST respectively). The deep learning-enhanced µ-maps (µ-mapDL-MLACF-PRE and µ-mapDL-MLACF-POST) were compared against MLACF-derived and CT-based maps (µ-mapCT). The performance of the method was also evaluated by assessing PET images reconstructed using each µ-map and computing volume-of-interest based standard uptake value measurements and percentage relative mean error (rME) and relative mean absolute error (rMAE) relative to CT-based method. No statistically significant difference was observed in rME values for µ-mapDL-MLACF-PRE and µ-mapDL-MLACF-POST both in fat-based and water-based soft tissue as well as bones, suggesting that presence of the radiopharmaceutical activity in the body had negligible effects on the resulting µ-maps. The rMAE values µ-mapDL-MLACF-POST were reduced by a factor of 3.3 in average compared to the rMAE of µ-mapMLACF-POST. Similarly, the average rMAE values of PET images reconstructed using µ-mapDL-MLACF-POST (PETDL-MLACF-POST) were 2.6 times smaller than the average rMAE values of PET images reconstructed using µ-mapMLACF-POST. The mean absolute errors in SUV values of PETDL-MLACF-POST compared to PETCT were less than 5

Aim:To develop and evaluate a new approach for spatially variant and tissue-dependent positron range (PR) correction (PRC) during the iterative PET image reconstruction.Materials and Methods:The PR distributions of three radionuclides (18F, 68Ga, and 124I) were simulated using the GATE (GEANT4) framework in different material compositions (lung, water, and bone). For every radionuclide, the uniform PR kernel was created by mapping the simulated 3D PR point cloud to a 3D matrix with its size defined by the maximum PR in lung (18F) or water (68Ga and 124I) and the PET voxel size. The spatially variant kernels were composed from the uniform PR kernels by analyzing the material composition of the surrounding medium for each voxel before implementation as tissue-dependent, point-spread functions into the iterative image reconstruction. The proposed PRC method was evaluated using the NEMA image quality phantom (18F, 68Ga, and 124I); two unique PR phantoms were scanned and evaluated following OSEM reconstruction with and without PRC using different metrics, such as contrast recovery, contrast-to-noise ratio, image noise and the resolution evaluated in terms of full width at half maximum (FWHM).Results:The effect of PRC on 18F-imaging was negligible. In contrast, PRC improved image contrast for the 10-mm sphere of the NEMA image quality phantom filled with 68Ga and 124I by 33 and 24%, respectively. While the effect of PRC was less noticeable for the larger spheres, contrast recovery still improved by 5%. The spatial resolution was improved by 26% for 124I (FWHM of 4.9 vs. 3.7 mm).Conclusion:For high energy positron-emitting radionuclides, the proposed PRC method helped recover image contrast with reduced noise levels and with improved spatial resolution. As such, the PRC approach proposed here can help improve the quality of PET data in clinical practice and research.

Aim To evaluate the effect of combining positron range correction (PRC) with point-spread-function (PSF) correction and to compare different methods of implementation into iterative image reconstruction for 124 I-PET imaging. Materials and methods Uniform PR blurring kernels of 124 I were generated using the GATE (GEANT4) framework in various material environments (lung, water, and bone) and matched to a 3D matrix. The kernels size was set to 11 × 11 × 11 based on the maximum PR in water and the voxel size of the PET system. PET image reconstruction was performed using the standard OSEM algorithm, OSEM with PRC implemented before the forward projection (OSEM+PRC simplified) and OSEM with PRC implemented in both forward- and back-projection steps (full implementation) (OSEM+PRC). Reconstructions were repeated with resolution recovery, point-spread function (PSF) included. The methods and kernel variation were validated using different phantoms filled with 124 I acquired on a Siemens mCT PET/CT system. The data was evaluated for contrast recovery and image noise. Results Contrast recovery improved by 2–10% and 4–37% with OSEM+PRC simplified and OSEM+PRC, respectively, depending on the sphere size of the NEMA IQ phantom. Including PSF in the reconstructions further improved contrast by 4–19% and 3–16% with the PSF+PRC simplified and PSF+PRC, respectively. The benefit of PRC was more pronounced within low-density material. OSEM-PRC and OSEM-PSF as well as OSEM-PSF+PRC in its full- and simplified implementation showed comparable noise and convergence. OSEM-PRC simplified showed comparably faster convergence but at the cost of increased image noise. Conclusions The combination of the PSF and PRC leads to increased contrast recovery with reduced image noise compared to stand-alone PSF or PRC reconstruction. For OSEM-PRC reconstructions, a full implementation in the reconstruction is necessary to handle image noise. For the combination of PRC with PSF, a simplified PRC implementation can be used to reduce reconstruction times.

Ziel/Aim To evaluate positron range (PR) corrections for iterative image reconstruction methods for I-124 PET imaging.