Mohammadreza Teimoorisichani

More than a decade has passed since the seminal work of Defrise et al (2012 Phys. Med. Biol. 57 885–99), which demonstrated that time-of-flight (TOF) PET can, under appropriate conditions, be intrinsically self-correcting for several physical degradation factors, most notably attenuation, which in contemporary clinical practice is typically derived from CT in PET/CT systems. The work of Defrise et al (2012 Phys. Med. Biol. 57 885–99) popularized the concept of maximum-likelihood reconstruction of attenuation and activity (MLAA) and stimulated extensive subsequent methodological and clinical investigations. The many studies that were done on this subject, led to early clinical evaluations of joint activity–attenuation reconstruction algorithms based on advanced optimization strategies. In recent years, however, self-attenuation correction has received comparatively less attention, as the field has increasingly shifted toward the synthesis of attenuation maps using deep learning (DL) often in combination with physics-based constraints. Many of the proposed DL-based methods worked with non–attenuation-corrected PET images, sometimes supplemented by additional sources of information. Within this paradigm, attenuation estimation has effectively become an image-to-image translation problem, well suited to convolutional neural networks, generative adversarial networks, and more recently emerging diffusion-based models. These approaches benefit from the widespread availability of PET/CT datasets, which enable the construction of large training cohorts, although the inherently sequential nature of PET and CT acquisitions can introduce misregistration and associated biases. Against this backdrop, the DL evolution raises a fundamental question: what level of attenuation information can truly be recovered from suboptimal, incomplete or noisy PET emission data alone? Addressing this question challenges not only the expressive power of data-driven methods, but also their physical interpretability and robustness. More broadly, it invites reflection on whether decades of physics-based modeling, analytical reconstruction theory, and detector-driven innovation will be complemented, or reinterpreted within increasingly more powerful machine learning frameworks. In this review, we examine these competing and often complementary approaches to CT-less quantitative PET imaging, with a particular emphasis on TOF-enabled methods, while recognizing that definitive answers to these questions may only emerge through continued theoretical, technological, and clinical investigation.

Flexible Magnetic Resonance (Imaging) (MR) hardware, such as body coils, can cause substantial attenuation of the Positron Emission Tomography (PET) signal in PET/MR imaging. Despite this, such hardware is typically not accounted for in clinical attenuation correction (AC), as its variable positioning and invisibility in standard MR sequences present significant challenges. In this study, we investigate the use of Maximum Likelihood Estimation of Attenuation and Activity (MLAA) guided registration of template hardware attenuation maps (μ-maps) to support AC for flexible hardware on PET data without Time Of Flight (TOF) information. Assuming the patient or phantom attenuation as known, attenuation maps beyond the patient outline were reconstructed from emission data and used as reference images for the non-rigid registration of pre-defined hardware template attenuation maps. The approach was evaluated and compared to the performance of MLAA-reconstructed attenuation alone on both phantom and patient data acquired on the Siemens Biograph mMR, including ¹⁸F-Fluorodeoxyglucose (¹⁸F-FDG) and ⁶⁸Ga-DOTATATE tracers, and at various count levels. For the phantom data, both MLAA-only and MLAA-guided registration of template hardware attenuation maps improved quantification. Registration-based correction generally resulted in closer agreement with the ground truth, with residual errors of up to ≈5% that were largely localized, while MLAA-only reconstructions showed deviations of up to ≈10%, tending to be more spatially extended. The method also tolerates moderate errors in MR-based AC and non-physical deformations of the hardware template, with quantification most sensitive to gross misplacement of the hardware. These findings suggest that MLAA-guided registration offers a practical and effective solution to address the current gap in attenuation correction for flexible hardware in clinical PET/MR imaging, also in non-TOF settings.

Purpose. Achieving high-quality PET imaging while minimizing scan time and patient radiation dose presents significant challenges, particularly in the absence of CT-based attenuation maps. Joint reconstruction algorithms, such as MLAA and MLACF, partially address these challenges but often result in noisy and less reliable images. Denoising these images is critical for enhancing diagnostic accuracy. Approach. This study introduces a novel cascaded relevancy-aware Generative Adversarial Network (reGAN) to improve the denoising and diagnostic reliability of μ-maps derived from joint reconstruction algorithms, ultimately aimed at enhancing PET imaging quality. The reGAN architecture employs a cascaded design incorporating UPlus GAN modules, relevancy mapping, and contextual attention mechanisms. The model was trained using PET/CT data from 16 patients, with MLAA and MLACF-derived μ-maps as input and CT-based μ-maps as the ground truth. Performance was evaluated using metrics such as SSIM, PSNR, VIF, and MSE. Comparative studies were conducted against other popular 2D and 3D GAN architectures. Results. The proposed reGAN achieved the highest SSIM (0.91 for MLAA and 0.93 for MLACF), PSNR (34.7 dB for MLAA and 36.2 dB for MLACF), and VIF (0.89 for MLAA and 0.91 for MLACF), while maintaining the lowest MSE (0.021 for MLAA and 0.018 for MLACF). Qualitative analysis demonstrated that reGAN preserved fine details, particularly in bony structures, and reduced artifacts effectively. Additionally, relevancy maps provided pixel-wise confidence indicators, further aiding interpretability and diagnostic reliability. Conclusion. reGAN presents a robust approach to medical image denoising, combining advanced generative modeling with diagnostic confidence metrics. The proposed method constitutes a viable approach for achieving quantitative accuracy in low-dose PET imaging in the absence of CT.

Single scatter simulation (SSS) is the most commonly used approach to mitigate the effects of scattered photons in PET imaging. However, the introduction of long axial field of view (LAFOV) scanners has required this method to be revisited or alternative methods to be adopted. In this work we compared a version of SSS with tail fitting (SSS-TF) devised for LAFOV, an alternative version of SSS with a statistical method to scale the scatter sinogram (SSS-MLSS), and a simplified fast GPU-based Monte-Carlo method (MC-GPU). These methods were evaluated using the MC toolkit GATE, simulating geometrical and realistic patient-based voxelized phantoms, employing the scatter simulated by GATE as reference. Furthermore, the three scatter correction methods were compared on scanned phantoms and patients. Regarding image artefacts, results showed that SSS-TF and SSS-MLSS performed similarly in general, but SSS-MLSS outperformed SSS-TF in challenging scenarios. MC-GPU consistently outperformed SSS-TF and SSS-MLSS regarding image artefacts, but at slightly longer computation times. Quantitatively, all methods showed relative differences <=5% compared to the reference, except for those regions with artefacts, but none of them showed consistently overall superior performance among them. Experimental measurements confirmed that SSS-MLSS outperforms SSS-TF in challenging cases, showing similar performance compared to MC-GPU.

Purpose: Long‐axial field‐of‐view (FOV) positron emission tomography (PET) scanners have gained a lot of interest in the recent years. Such scanners provide increased sensitivity and enable unique imaging opportunities that were not previously feasible. Benefiting from the high sensitivity of a long‐axial FOV PET scanner, we studied a computed tomography (CT)–less reconstruction algorithm for the Siemens Biograph Vision Quadra with an axial FOV of 106 cm.Methods: In this work, the background radiation from radioisotope lutetium‐176 in the scintillators was used to create an initial estimate of the attenuation maps. Then, joint activity and attenuation reconstruction algorithms were used to create an improved attenuation map of the object. The final attenuation maps were then used to reconstruct quantitative PET images, which were compared against CT‐based PET images. The proposed method was evaluated on data from three patients who underwent a flurodeoxyglucouse PET scan.Results: Segmentation of the PET images of the three studied patients showed an average quantitative error of 6.5%–8.3% across all studied organs when using attenuation maps from maximum likelihood estimation of attenuation and activity and 5.3%–6.6% when using attenuation maps from maximum likelihood estimation of activity and attenuation correction coefficients.Conclusions: Benefiting from the background radiation of lutetium‐based scintillators, a quantitative CT‐less PET imaging technique was evaluated in this work.

Objective . The aim of the phantom study was to validate and to improve the computed tomography (CT) images used for the dose computation in proton therapy. It was tested, if the joint reconstruction of activity and attenuation images of time-of-flight PET (ToF-PET) scans could improve the estimation of the proton stopping-power. Approach . The attenuation images, i.e. CT images with 511 keV gamma-rays ( γ CTs), were jointly reconstructed with activity maps from ToF-PET scans. The β ⁺ activity was produced with FDG and in a separate experiment with proton-induced radioactivation. The phantoms contained slabs of tissue substitutes. The use of the γ CTs for the prediction of the beam stopping in proton therapy was based on a linear relationship between the γ -ray attenuation, the electron density, and the stopping-power of fast protons. Main results . The FDG based experiment showed sufficient linearity to detect a bias of bony tissue in the heuristic look-up table, which maps between x-ray CT images and proton stopping-power. γ CTs can be used for dose computation, if the electron density of one type of tissue is provided as a scaling factor. A possible limitation is imposed by the spatial resolution, which is inferior by a factor of 2.5 compared to the one of the x-ray CT. γ CTs can also be derived from off-line, ToF-PET scans subsequent to the application of a proton field with a hypofractionated dose level. Significance. γ CTs are a viable tool to support the estimation of proton stopping with radiotracer-based ToF-PET data from diagnosis or staging. This could be of higher potential relevance in MRI-guided proton therapy. γ CTs could form an alternative approach to make use of in-beam or off-line PET scans of proton-induced β ⁺ activity with possible clinical limitations due to the low number of coincidence counts.

In recent years, there has been a renewed interest in brain-dedicated PET imaging systems, particularly in the context of combined PET/MR imaging. We are currently designing a brain-dedicated PET insert suitable for an ultra-high field brain-dedicated MR scanner, the Siemens Magnetom 7T MR scanner. In this paper, an investigation on the count rate performance of several possible detectors through a series of Monte Carlo simulations is reported. Brain-dedicated PET scanners with a lutetium oxyorthosilicate scintillator and a detector area of 0.04 (1 crystal per detector) to 101.37 (2500 crystals per detector) cm ² , detector thickness of 10 to 20 mm and a fixed crystal pitch of ~2 mm were simulated. The count rate performance of each scanner was evaluated as a function of detector deadtime type and constant, coincidence timing window and lower level discriminator. Also, the effects of activity outside the field-of-view (FOV) on the count rate performance of each scanner were studied. For each detector geometry and performance metric, the scanner singles rate, scanner sensitivity and noise equivalent count rate as a function of activity in the FOV were measured. It was seen that scanners with detectors comprised a few crystal elements showed reduced scanner sensitivity due to a high number of inter-detector scattering. The count rate performance of scanners with large detectors, on the other hand, was mainly determined by the deadtime properties of the detectors. A model for the count rate performance of the scanner with each studied detector is presented in this work.

A dual-layer offset (DLO) detector enables depth-of-interaction (DOI) through light sharing between two layers of scintillation arrays with a single-ended readout (SER) scheme. However, the SER scheme in DLO detectors may lead to a layer misassignment when inter-crystal scattering occurs. The aim of this work is to study inter-crystal scattering and evaluate the effects of layer misidentifications in DLO detectors on the performance of scanners suitable for a brain-dedicated PET insert. The influence of layer misidentification on the coincidence response functions (CRFs) of 3 different DLO detectors with total/front/back layer thicknesses of 15/6/9 mm, 20/8/12 mm, and 25/7.5/17.5 mm and a crystal width of about 3 mm was studied through Monte Carlo simulations. To overcome layer misidentification, we studied a practical DLO detector design in which each layer can be read out independently through a discrete-layer readout (DLR) scheme where light sharing between the layers is avoided. The CRFs of the mentioned DLO detectors assuming SER and DLR were analyzed. To evaluate the effects of layer misidentification on image quality, images of a Derenzo-like phantom were also reconstructed for all DLO and their equivalent single layer PET scanners. Our analysis showed that layer misassignments due to inter-crystal scatter in DLO detectors mainly has effect on the full-width at tenth maximum of the CRFs. According to the reconstructed images of the phantom, no significant improvements in the quality of the images were seen when SER was replaced with DLR. The results suggest that layer misidentification in DLO detectors does not play an important role in the quality of the PET images.

PET scanners suffer from a loss of resolution due to parallax error. Depth-of-interaction (DOI) detectors, such as dual-layer offset (DLO) detectors, mitigate the loss of resolution. We are designing a DLO brain dedicated PET insert that fits into the Siemens Magnetom 7T brain MR scanner. A wide range of DLO detector geometries was evaluated through Monte Carlo simulations in this work. The total detector thickness was varied from 10 to 30 mm and for each total thickness, various layer thickness ratios and crystal widths in the range of 0.5 to 4 mm were examined. The effects of each detector geometry on radial mispositioning of the incident photons, detection efficiency ratio between the two layers, coincidence response functions (CRFs), system spatial resolution, and sensitivity were studied. Optimum layer thickness ratio for each total thickness is discussed. Analysis of CRFs showed that the improvement in the CRF of a DLO PET scanner with a total detector thickness of larger than 10 mm is lower for detectors with a crystal width of smaller than 2 mm than for detectors with a larger crystal width. Estimated spatial resolution and sensitivity of a PET scanner with each suggested detector geometry are also reported.