Realization of long axial field-of-view PET technology in the clinic and research environment

Abstract

Long axial field-of-view (AFOV) PET-CT instruments have significantly higher sensitivity than conventional PET scanners, allowing for reduced scan time and/or reduced radiotracer dose. The long axial coverage also permits multiorgan dynamic imaging with fine temporal sampling. In this article, we discuss how the technical advantages of long AFOV PET instruments are being realized in the clinic and exploited in research studies. We describe some practical considerations for these systems in the clinic, as well as their benefits beyond the clinic. We consider alternate scanner designs to mitigate the high cost of long AFOV systems currently in use and describe how deep learning is being applied to long AFOV PET studies to reduce image noise, allowing for even shorter scans to minimize motion, aid in dynamic imaging applications, and image generation. We conclude with some ideas on what future developments may be on the horizon for long AFOV PET.

Introduction

Long axial field-of-view (AFOV) PET-CT systems with an AFOV of >60 cm were first constructed in 2018, years after the theoretical benefits of such systems were first discussed^1–7 and earlier attempts were made to develop prototype scanners.^8,9 These systems have significantly higher sensitivity than conventional PET systems with an AFOV of 15-35 cm. This technical advantage can be leveraged for reduced scan time and/or reduced dose, as well as fine temporal sampling in multiorgan dynamic studies.^10–15 It has been 7 years since the first human scans were acquired on long AFOV PET systems.^16,17 Since that time, the number of commercial long AFOV PET systems in routine clinical/research use has expanded to nearly 100 worldwide. The broad interest in total body (TB) or long AFOV PET imaging is underscored by the number of dedicated workshops and sessions at major scientific conferences that have focused on various aspects of long AFOV PET, as well as a rapidly growing number of review articles on this topic.^18–20

In this article, we discuss how the technical advantages of long AFOV PET are being realized in the clinic and exploited for novel research studies. We cover the state-of-the-art instrumentation, developments in the use of deep learning (DL) methods for image generation, and illustration of some applications promoted at the University of Pennsylvania (UPenn).

Existing systems

There are currently four models of long AFOV PET systems in routine operation, summarized in Table 1.^21–24 All of these systems are 3D with time-of-flight (TOF) capability and are based on technology available in conventional PET-CT scanners, specifically lutetium-based (LSO, LYSO) detectors with silicon photomultiplier (SiPM) digital readout. Thus, these systems maintain the intrinsic performance achieved on conventional scanners (e.g., spatial, energy, and TOF resolutions) while dramatically increasing the sensitivity by extending the AFOV with a large axial acceptance angle. A range of AFOVs is available, from the Siemens Quadra with 106 cm, covering the head to thighs, to the United Imaging uEXPLORER with 194 cm, covering the entire body of most adults. All systems have a ring diameter of 76-78 cm, the same as for conventional PET-CT scanners.

Table 1.

Long AFOV PET systems in routine use.

System	United Imaging uEXPLORER	United Imaging uMI Panorama GS	PennPET Explorer	Siemens Biograph Vision Quadraa
Crystal size (mm3)	2.76 × 2.76 × 18.1	2.76 × 2.76 × 18.1	3.86 × 3.86 × 19	3.2 × 3.2 × 20
Crystals: SiPM	10.5:1	9:4	1:1	5:1
AFOV (cm)	194	148	142	106
Axial acceptance angle	±57°	±62°	±62°	±18°/±52° (MRD 85/322)
TOF resolution (ps)	505	189	250	225/230 (MRD 85/322)
Spatial resolution @center
Transverse (mm)	3.0	2.9	4.0	3.3
Axial (mm)	2.8	2.8	4.0	3.8 (MRD 85)

^aFor the Siemens Biograph Vision Quadra, performance metrics are reported for its two maximum ring difference (MRD) settings: high sensitivity (HS) with an MRD of 85 and ultra-high sensitivity (UHS) with an MRD of 322.

The total sensitivity gain of a long AFOV system compared with a standard AFOV system is ∼15-40x¹², but the peak (e.g., per-organ) sensitivity gain is ∼2-3x because oblique coincidences are attenuated more due to the longer path length in the body. Thus, much of the peak sensitivity benefit can be achieved with an AFOV <1 m, as was found with the prototype PennPET Explorer with a 64-cm AFOV.²⁵ Such a moderate AFOV leads to high-quality images of a single organ, for example, the brain or heart, and may also be sufficient for clinical static imaging from eyes-to-thighs with FDG, with multibed scanning used for taller patients—but with a considerably shorter scan duration compared to conventional systems. A longer AFOV, of course, has peak sensitivity covering more organs, in addition to higher total sensitivity overall, and allows high-temporal-resolution dynamic imaging throughout the body, including a vascular structure, for an image-derived input function (IF).^26,27 An additional advantage of long axial coverage is the availability of large blood pools that reduce the impact of partial volume effects on the IF.

While the existing long AFOV PET systems maintain the intrinsic performance of conventional scanners, their increased length and larger axial acceptance angle could be anticipated to impact overall imaging performance. There is a small (∼0.5 mm) degradation in the axial spatial resolution at the center of the AFOV due to axial parallax errors,^23,24 but this has no impact on the measured contrast recovery coefficient of even 10-mm spheres as measured on the PennPET Explorer.²³ The scatter fraction is similarly not impacted by the increased number of detected events,^22,23,25 confirming predictions of early simulation studies.¹ The randoms fraction increases for the longer systems due to the higher singles rates but is still manageable (e.g., increasing by a factor of ∼1.3 on the PennPET Explorer going from a 64-cm AFOV to 142 cm^23,25) also as predicted by simulation.¹ The TOF resolution degrades by <10 ps in going from a restricted to wide-open axial acceptance angle.^22,23 Consequently, the impact of extending the AFOV on imaging performance is seen primarily in the large increase in sensitivity.

Considerations with long AFOV systems

Choice of axial acceptance angle

A key parameter that determines the sensitivity gain of a 3D system is the axial acceptance angle. A very large amount of data is acquired on long AFOV PET systems with open axial acceptance angles that capture coincidences over the entire volume of the scanner. This results in increased data storage and computational demands with longer reconstruction times for fully corrected images. This can be decreased by restricting the axial acceptance angle, but at the risk of increased image noise due to the loss of oblique true events.²⁸

Different approaches have been implemented with regard to restricting the axial acceptance angle. The PennPET Explorer operates with the full axial acceptance angle (62°); the uEXPLORER has a restricted angle (57°), corresponding to coincidences between five of its eight total rings, while the Panorama, with a similar AFOV to the PennPET Explorer, uses a comparable acceptance angle (62°).²⁴ The Siemens Quadra offers two modes of acquisition: (i) a high sensitivity (HS) mode with a maximum ring difference (MRD) of 85 and an acceptance angle (18°) equal to that of the Siemens Vision (26.3-cm AFOV), and (ii) an ultra-high sensitivity (UHS) mode with a MRD of 322 (52°), allowing all possible cross-coincidences. At UPenn, the PennPET Explorer is used primarily for research, whereas the Siemens Quadra is currently used primarily for clinical imaging, with nuclear medicine physicians receiving both the HS and UHS images for interpretation. Although the UHS mode offers higher statistical quality, clinicians currently rely on HS images to facilitate comparisons across scanners. We anticipate that as clinicians become more familiar with the Quadra data, the use of the UHS images will become more routine, recognizing that newly visualized small foci (which may not be visible on HS images) may reflect improved imaging characteristics rather than new disease.²⁹

Variations in sensitivity across the AFOV, as well as variations in image noise, are mitigated by attenuation effects of oblique lines-of-response.¹⁵ As an example, Figure 1A shows the pixel-to-pixel image noise as a function of axial position in a uniform 10-cm diameter, long pipe phantom for several axial acceptance angles. The image noise is uniform across much of the AFOV and lower with a wider acceptance angle. Figure 1B compares the image noise near the center of the AFOV (where the axial acceptance angle is widest) for several human subjects (body mass index, BMI = 24-51 kg/m²) measured in the liver. Increasing the axial acceptance angle reduces the noise in the center of the AFOV, although this is more significant for the phantom with the smallest cross-sectional size. For patients, attenuation of oblique events leads to relatively little improvement in image noise for angles larger than 40°, suggesting that the axial acceptance angle can be restricted for patients with average to large BMIs.

Figure 1.

(A) Axial noise profile measured on the PennPET Explorer in a 10-cm diameter, long uniform pipe phantom (75 MBq, 3-min acquisition) for several axial acceptance angles from 17°, corresponding to a single ring, to 62°, corresponding to a wide-open acceptance angle. (B) Change in image noise near the axial center for the pipe phantom and several patients (FDG) compared with the noise for a 62° axial acceptance angle.

Ultimately, the choice of acceptance angle will depend on the intended application and need for faster data processing. In general, a restricted axial acceptance angle may not measurably impact image quality for standard of care (SOC) FDG studies with a standard dose, whereas a larger angle may be preferred for studies involving lower injected doses and/or shorter scan durations, pediatric studies (with lower attenuation), or radionuclides with low positron branching ratio, such as ⁶⁴Cu or ⁸⁹Zr.

Patient positioning

A practical consideration for imaging on a long AFOV PET system is positioning the patient. Peak sensitivity occurs at the axial center of the scanner. However, attenuation effects in patients result in relatively uniform sensitivity across the AFOV, except near the axial ends. Consequently, for whole-body studies, exact patient positioning is typically not critical. On the PennPET Explorer with 142-cm AFOV, patients are positioned to allow for imaging from the top of the brain to mid-thigh. For single-organ imaging (e.g., brain, heart), this default placement may position the organ of interest away from the region of highest sensitivity. Thus, for organ-specific studies on the PennPET Explorer, patients are typically positioned with the organ of interest near the axial center to take full advantage of the system’s maximum sensitivity. The longer tunnel may cause a modest increase in claustrophobia for some patients, which can be mitigated by positioning their head closer to the end of the scanner, as well as by reducing the scan duration.

A further consideration for single-organ imaging (e.g., cardiac or brain) performed on a long AFOV PET system is that a full-length CT scan may not be necessary. For whole-body studies, full CT coverage provides important anatomic context but adds unnecessary radiation when imaging a single target organ that occupies only a small portion of the AFOV. DL-based methods can generate surrogate mu-maps for accurate quantitative analysis (Figure 2) and help to avoid spatial mismatches between the CT and PET due to patient or organ motion. Recent work^30,31 has shown that hybrid attenuation correction strategies, such as extending short, localized CT scans using scout-derived synthetic CT or deep-learning-based full-body CT extrapolation, can provide accurate mu-maps for attenuation correction on long AFOV systems without requiring full-CT coverage.

Figure 2.

DL-based attenuation correction results from UPenn. (A) Brain study using [¹⁸F]-FDG and an [¹¹C]-labeled tau tracer (the latter not included in network training). Transmission (Tr) images generated from nonattenuation corrected (NAC) PET using the DL method are compared with the reference CT-derived Tr images. For both tracers, Tr image values agree within 1%. For the FDG study, the grey-to-white matter ratios measured on the PET images were 3.80 for CT-AC and 3.77 for DL-AC. (B) Cardiac example comparing Tr images from CT and DL methods, along with PET images reconstructed using CT-AC and DL-AC, showing minimal differences. The heart SUVmean values in a volume of interest (VOI) were 3.76 for CT-AC PET and 3.81 for DL-AC PET. Data courtesy of J. Dubroff, I. Nasrallah, and C. Wiers, UPenn.

Reconstruction times for long AFOV PET data

Long AFOV PET imaging introduces significant data management challenges, as noted in other review articles.^15,32,33 The much larger datasets, particularly those generated for dynamic imaging and kinetic analysis, demand greater computational efficiency and sophisticated modeling methodologies. These massive data volumes, driven by high coincidence rates and the added temporal dimension (in case of dynamic imaging), lead to longer reconstruction times for fully corrected images. Thus, nonquantitative images are often generated to ensure a successful scan before the patient leaves the clinic. On the PennPET Explorer, for example, an uncorrected single-view (TOF deposited) histo-image³⁴ can be generated in seconds and reviewed on demand for quality control purposes.

DL-based reconstruction frameworks can also be utilized to address the computational demands of image reconstruction from these large datasets. FastPET^35,36 is one such solution. These frameworks replace slow iterative algorithms with a fast end-to-end neural network, achieving 20-30x faster reconstructions than usual iterative methods. FastPET converts raw TOF list-mode data into multiview histo-images using most likely positioning based on the TOF information³⁴ and, together with CT-based attenuation maps, uses a 3D convolutional neural network to generate high-quality, denoised emission images directly from the data.

Benefits of long AFOV PET in the clinic

The benefits of the high sensitivity due to long AFOV PET imaging have been widely demonstrated, permitting high throughput with reduced injected doses while maintaining, or even improving, image quality.^37,38 Alberts et al³⁹ demonstrated equivalence between images acquired on the Vision and the Quadra (HS mode), but at 1/10th of the time on the Quadra, correlating well with predictions based on sensitivity differences.¹⁵ Other studies on the Quadra⁴⁰ and uEXPLORER⁴¹ have demonstrated improved detectability and staging in small tumors compared with conventional systems, thereby impacting clinical management. At UPenn, with the Siemens Quadra, we have reduced the injected FDG activity from 370 to 296 MBq and image for 5 min from the skull vertex (as opposed to the skull base) through the thigh. While a further reduction in dose or scan duration is possible given the high quality of the images, practical constraints of available technologists and radiologists limit our throughput to 3-4 patients per hour. We also note that the time for patient positioning is independent of the scanner and is often the rate-limiting step. Of course, other sites have dramatically increased patient throughput. In a study from China with the uEXPLORER, 60 patients were scanned in 8 h, with an average PET imaging time of 3 min with 8 min spent in the scanning room.⁴² In addition to higher throughput to recover the higher capital cost, faster scans have other benefits. Shorter scan durations reduce the likelihood of gross motion and discomfort due to claustrophobia. It can obviate the need for sedation in pediatric subjects^43–45 or patients with dementia.

Delayed imaging of FDG has also been implemented in clinical studies, where imaging happens 2-3 h post-injection, instead of 1 h, to allow for clearance of tracer from normal tissue/blood pool.^20,46,47 At UPenn, for example, vasculitis is imaged with a 2-h uptake time after the injection of 300 MBq FDG.⁴⁸ Long AFOV PET systems have also facilitated imaging of nonstandard radionuclides with low branching ratios for positron emission or long half-lives that limit the injected activity to restrict the patient’s radiation dose (e.g., ⁹⁰Y, ⁸⁹Zr). On the PennPET Explorer, an injected ⁸⁹Zr dose of 37 MBq yields good image quality for a 30-min scan duration when imaged 24 h or longer post-injection. Robust DL-based denoising methods can further enhance image reliability and quantitative performance for these tracers (Figure 3), achieving good image quality with scan durations under 5 min. This optimization is critical for the integration of such protocols into routine clinical workflows to expand their clinical utility in therapy planning and treatment monitoring.

Figure 3.

Imaging study on the PennPET Explorer with a ⁸⁹Zr-labeled tracer targeting CD8+ T-cells. A research subject (BMI 26.6 kg/m²) was scanned for 30 min at 24 h after injection of 41 MBq of an ⁸⁹Zr-labeled anti-CD8 minibody (crefmirlimab berdoxam). The top row shows the original 30-min PET scan alongside images retrospectively sub-sampled to 15-, 10-, 5-, and 2-min acquisition times. The bottom row presents the corresponding images following deep-learning-based denoising. Data courtesy of M. Farwell, UPenn, and with acknowledgment to ImaginAb for providing the compound.

Benefits of long AFOV PET beyond the clinic

Dynamic imaging

The emergence of long AFOV PET scanners has expanded the potential of whole-body dynamic imaging to add diagnostic value beyond clinical SOC imaging. For example, for oncological applications, dynamic PET enables quantification of tissue-tracer interactions, improving the distinction between specific (pathological) and nonspecific (physiological) uptake in several cancer types.^49–53 With conventional AFOV PET scanners, high-temporal resolution is limited to a short axial range, restricting their ability to capture rapidly changing or early kinetics for more than a single organ.⁵⁴ In contrast, the high sensitivity and axial coverage of long AFOV systems allow list-mode data to be reframed into very short time intervals (<5 s) for the full body, supporting advanced kinetic analyses and parametric imaging.^55,56

Besides single-tracer imaging studies, an emerging approach (leveraging the sensitivity gains of a longer AFOV) involves the use of two radiotracers in a single imaging session to quantify different physiological pathways, providing multiplexed information for malignancy characterization. Several groups have implemented dual-tracer dynamic imaging protocols, where an acquisition (often abbreviated) of the first radiotracer, injected at low dose, is followed by an injection of a higher dose of the second radiotracer and subsequent dynamic imaging. Examples of such dual-tracer protocols include [⁶⁸Ga]-PSMA + FDG^57,58 or [⁶⁸Ga]-FAPI + FDG,^59,60 in addition to [¹⁸F]-FGln + FDG for breast cancer imaging.⁶¹ As illustrated in Figure 4 for the example of an FGln/FDG dual-tracer study, the high sensitivity of the long AFOV PET allows for the FGln study with a lower dose to have sufficient image quality for accurate kinetic modeling of the glutamine volume of distribution of the lesion, without impacting the diagnostic interpretation of FDG uptake.

Figure 4.

A dual-tracer protocol of a breast cancer patient scanned on the PennPET Explorer for 30 min with a dose of 35 MBq [¹⁸F]Fluoroglutamine (FGln), followed by a dose of 290 MBq of Fluorodeoxyglucose (FDG) injected at 30 min with an additional 60 min scan.

Total-body perfusion imaging

Another exciting example is in total-body perfusion measurements, where knowledge of tissue perfusion throughout the body is important in disease management.^62–64 Recent work⁶⁵ also demonstrates how the gain in temporal resolution directly expands what can be achieved with widely available tracers such as FDG. Historically, blood flow imaging has been limited to tracers with high extraction fraction (e.g., [¹⁵O]-H₂O) because conventional AFOV PET systems cannot resolve the rapid vascular transit that occurs within the first seconds after injection. With a long AFOV PET scanner, the early vascular phase of FDG can now be reconstructed at 1-2 s/frame, allowing the rapid bolus passage to be captured with sufficient fidelity to enable direct estimation of blood flow (e.g., K1) rather than relying on surrogate parameters such as standardized uptake value (SUV).

Multiorgan interaction quantification

Another emerging use of long AFOV systems is in multiorgan interactions (e.g., brain/heart⁶⁶) in disease or with interventions (e.g., rest/stress paradigm⁶³) Having the ability to image the entire body dynamically with fine temporal sampling has enabled research in this area to flourish. The PennPET Explorer is in a research environment where studies of cardiac/kidney function and cardiac/brain interactions are performed in large animals⁶⁷ prior to translation to humans. For example, at UPenn, studies of mu-opioid receptors in the central nervous system (CNS) and in the periphery under baseline and blocking conditions were first performed in nonhuman primates,⁶⁸ and then, translated to humans.⁶⁹ A major advantage of evaluating extra-CNS receptor binding is to establish a reference region for simplified modeling.⁶⁹ Further, multiorgan evaluations can be performed to assess receptor occupancy with different blocking agents, where there is differential tracer uptake depending on the binding characteristics of the blocking agent (e.g., binding affinity, P-glycoprotein substrate).⁷⁰

Biodistribution studies

While most studies involving long AFOV PET are related to clinical applications or kinetic modeling for disease characterization, these systems have also found important applications in radiopharmaceutical development and testing, where the high sensitivity and extended axial coverage become particularly relevant.^71,72 The high sensitivity allows low-dose injections of candidate radiotracers without requiring full scale-up of radiochemical production prior to verification of the radiotracer’s utility. It permits delayed imaging of these candidate radiotracers over several half-lives for biodistribution and radiation dosimetry studies. Similarly, for radionuclide therapy, there is a need to measure the dosimetry of these agents on an individual level. With conventional AFOV scanners, it can be challenging to capture the full-time course of the radioligands due to the poor statistics, whereas the higher sensitivity of long AFOV systems allows quantitative imaging over a longer period to capture the changing biodistribution for more accurate dosimetry assessment.

Future directions

Alternative scanner designs to mitigate cost

The cost of long AFOV PET/CT systems is >3 times that of conventional PET/CT scanners. Although a cost analysis with emphasis on clinical throughput has demonstrated the benefit of a long AFOV system over 1-2 standard AFOV scanners,⁷³ the higher cost may be prohibitive for many sites. There are a number of proposed designs to lower the cost by reducing the number of detectors while maintaining a sensitivity gain over conventional AFOV systems and a larger axial coverage to allow for dynamic multiorgan scanning.⁷⁴ For PennPET Explorer, each ring operates independently, acquiring the data as a “singles” data stream and concurrently generating coincidence list data in software.²³ Such a design is configurable and allows the user a choice of AFOV depending on their needs and budget. In fact, the PennPET Explorer has operated for human imaging in multiple configurations ranging from three rings (64 cm) up to six rings (142 cm). Similarly, GE Healthcare has recently offered a configurable version of their Omni PET-CT (32-128 cm AFOV). The PennPET Explorer has also operated with gaps between rings. Figure 5 shows an example of the same patient imaged the first time with five rings with inter-ring gaps (7.6-cm gaps between 16.4-cm wide rings) and later on the current configuration with six rings (no gaps, 22.9-cm wide rings), demonstrating excellent image quality in both configurations—one with 60% of the detector cost of the other. Various designs of PET scanners with sparsity levels close to 50% have been studied^75–77 with recent interest in cost-effective designs for long AFOV PET.^78–80 These designs achieve large axial coverage with fewer detectors by trading off sensitivity, but deep learning denoising or sensitivity-weighted reconstructions^81,82 may help offset the sensitivity loss by mitigating noise and nonuniformities introduced by sparse geometries. Other low-cost design concepts for long AFOV PET imaging that are under development include the Walk-Through PET^83,84 and the IMAS system,⁸⁵ as well as designs to use less expensive detector materials (e.g., BGO)⁸⁶ or the J-PET system with plastic scintillators.⁸⁷

Figure 5.

(A) Prototype PennPET Explorer: five rings (112-cm AFOV) with axial gaps. (B) PennPET Explorer: six rings (142-cm AFOV) with full 7-row readout. The axially sparse 5-ring design achieves excellent image quality with large axial coverage at ∼60% of the system cost (25 vs 42 detector rows). Data courtesy of C. Wiers, UPenn.

Emerging imaging capabilities

Positronium lifetime imaging

Long AFOV PET systems have also been used to perform feasibility studies of novel imaging methods that may not be possible on conventional PET scanners. This includes positronium lifetime imaging (PLI) to probe the molecular environment in tissue, first demonstrated by Moskal et al⁸⁸ in humans by imaging a glioblastoma brain tumor with [⁶⁸Ga]-DOTA-SP. Utilizing a high-resolution reconstruction method developed by Huang and Qi⁸⁹ and leveraging the higher performance of the PennPET Explorer compared to the J-PET scanner, we have illustrated the feasibility of high-resolution PLI using ⁸²Rb, ⁶⁸Ga, and ⁴⁴Sc with consistent accuracy across varying activity levels.⁹⁰ The high sensitivity of long AFOV PET facilitates measurement of triple coincidences (two 511-keV gammas with a prompt gamma) required for PLI and can be used to establish the clinical potential of this new technique.

High-resolution brain-body imaging

The long AFOV PET instruments from United Imaging and Siemens both have excellent spatial resolution; however, brain-body research applications may benefit from further improvements in spatial resolution in the brain. While a long AFOV PET system with higher-resolution detectors is achievable, motion and noncollinearity effects limit the resolution practically achieved in the body. Rather than driving up the cost of the full instrument with smaller crystals,⁹¹ one might add a pair of high-resolution “outsert” detectors⁹² to focus on (or magnify) regions of interest in the brain. Alternatively, we propose coupling a full high-resolution brain imaging with a long AFOV PET system as an additional ring to acquire cross-coincidences to permit imaging the brain and whole body simultaneously with maximum performance and without compromise. Such a combined system would serve a range of applications in oncology and neurology using an emerging selection of radiotracers to study disease with dynamic imaging.

Spectral CT/long AFOV PET

The PennPET Explorer is coupled with a spectral CT scanner (Philips IQon CT), thereby providing CT advances of material-specific x-ray attenuation that enable more quantitative evaluations, metal artifact reduction, better contrast conspicuities, and improvement of tissue characterization.⁹³ Market trends indicate that spectral CT scanners will soon succeed conventional CT scanners, and this integration provides the opportunity to develop and investigate innovative protocols that synergize these two modalities,⁹⁴ with the eventual goal of enabling new clinical applications not possible with each instrument alone. Potentially the combination of tissue perfusion from contrast-enhanced spectral CT with measures of metabolism from PET (with improved diagnostic quality enabled by long AFOV systems) will lead to clinically accessible protocols with additional quantitative metrics.

Application of deep learning

While the technical advancements of long AFOV PET systems open up new clinical and research opportunities for PET imaging, they also introduce significant challenges in data management, image processing, and interpretation. Artificial intelligence (AI) and deep learning (DL) are particularly well-positioned to address these challenges.^33,95–100 Broadly, DL applications in long AFOV PET can be categorized as: (i) extensions of methods already studied for conventional AFOV PET, now enhanced by richer spatial and temporal information from long AFOV PET data, and (ii) solutions tailored to challenges specific to long AFOV PET, such as handling large dynamic datasets or leveraging whole-body kinetics.

While many DL/AI techniques apply to both conventional and long AFOV systems, differences in sensitivity and axial coverage necessitate retraining with new data from long AFOV systems. The larger image volumes introduce extra challenges, including a wider dynamic range of tissue types, but at the same time, this highlights a key strength of long AFOV PET data for DL development: its ability to provide more diverse training data. This not only strengthens neural network training but can also facilitate translation of DL-based methods to shorter AFOV PET scanners through improved supervision (target) quality. Several studies have shown the successful transfer of neural networks trained on long AFOV to short AFOV PET.^101–103 For tasks like denoising, the higher sensitivity of long AFOV PET enables the routine acquisition of high-count, low-noise “ground truth” images within clinically feasible scan times, providing superior supervision targets for training networks. In the example of Figure 6, the deep learning denoising (DL-DN) model was trained on 10-min high-count data (50-60 mins p.i. FDG) as the target reference, with a range of subsampled scan durations (low-count data) as inputs. For testing, lesion detectability was assessed using 3-min data as reference, leveraging the higher signal-to-noise ratio (SNR) of the 10-min data for training while basing performance evaluation on a clinically realistic 3-min protocol. As shown in Figure 6, the DL-DN model achieved effective noise reduction across all subsampled inputs, and the resulting lower background noise in the liver and lung yielded substantial increases in lesion detectability, quantified by the area under the localization receiver operating characteristic (ALROC) curve.¹⁰⁴

Figure 6.

Visual comparison of 2-mm-thick coronal image slices with 10-mm spherical lesions embedded into the liver and lung at a 3:1 uptake ratio. The 3-min reference and low-count images (obtained by list-mode subsampling) are compared to the corresponding DL-DN images. Results from a numerical observer study using lesion-embedded patient data are shown, where lesion detectability was quantified using the generalized scan statistics methodology developed by Surti and Karp.¹⁰⁴ The ALROC curves for liver and lung lesions are plotted as a function of scan duration and compared between low-count and DL-DN. Data courtesy of C. Wiers, UPenn.

There has been increasing interest in CT-less attenuation and scatter correction (ASC) using DL to enable quantitatively accurate PET imaging without reliance on a CT.¹⁰⁵ This interest has continued with the advent of long AFOV PET systems, whose enhanced sensitivity allows for substantial dose reductions, thereby increasing the relative contribution of the CT component to overall radiation exposure. DL-based ASC approaches reported in the literature can be broadly categorized into two classes, both typically using noncorrected PET data as input. Indirect methods aim to predict a pseudo-CT image that can be used as a surrogate in the reconstruction process,^106,107 whereas direct methods seek to generate fully corrected PET images, requiring the neural network to implicitly learn both attenuation and scatter compensation.^108,109 These DL-based methods have the advantage of correct alignment of the attenuation and scatter corrections with the emission data in the presence of any motion throughout the AFOV, unlike a CT. While both approaches have demonstrated promising performance, most studies have been limited to single-tracer applications and/or single-organ imaging (brain) within conventional AFOV PET settings. Given the larger anatomical coverage of long AFOV PET, evaluating the applicability of these DL-ASC approaches across multiple tissues/organs (total body) and multiple tracer studies is important.^110–113 In parallel, hybrid strategies that integrate DL with established methods, such as maximum likelihood reconstruction of attenuation and activity or lutetium background-based techniques, have been proposed.^114,115

Applications of DL specific to long AFOV PET imaging address challenges unique to their extended coverage, particularly in dynamic multiorgan imaging. Examples include studies exploring DL for improved parametric imaging, advanced kinetic modeling, or abbreviated dynamic protocols, with the aim to make sophisticated quantitative analyses more clinically feasible.^33,98 Short time frames (<10 s) used for kinetic analysis remain susceptible to low-count noise, degrading both the IF estimation and time-activity curve (TAC) fitting.⁵¹ To mitigate this, DL-based denoising methods have been proposed to stabilize TACs and reduce variance in parametric maps. Examples include frame-wise image-based denoising networks,^116–118 spatio-temporal models,^119,120 and end-to-end networks that map dynamic emission data directly to parametric images.^121–123 DL-based protocol development is also underway. Although long AFOV PET systems make dynamic whole-body imaging more feasible, dynamic protocols remain complex and time-consuming, which limits their adoption in routine clinical workflows. The high sensitivity, however, creates the possibility of abbreviated dynamic protocols, for example, 15-20 min acquisitions^124,125 to measure kinetic parameters that routinely require 60 min. Proposed solutions include hybrid population-based IF^126,127 and AI-assisted frameworks that infer missing temporal information.^128–130 One area where these methods may particularly be valuable is dual-tracer imaging, as illustrated in Figure 3. Applying a DL-denoising network to a study that used a low [¹⁸F]-FGln dose (40 MBq) followed by a higher [¹⁸F]-FDG dose (400 MBq) showed that both doses could be reduced by an additional factor of 10, while still maintaining accurate kinetic parameter estimation.¹¹⁸ Future work may extend such approaches to tracer pairs with less favorable kinetics and enable similar low-high-dose dual-tracer protocols on standard AFOV PET systems that lack the sensitivity gains of long AFOV scanners.

Conclusions

Widespread use of the first generation of long AFOV PET systems over the last 5 years has demonstrated a significant upside of improved imaging performance compared with conventional PET systems and without any noticeable downside other than processing time and data sizes. Current long AFOV PET systems have established that AFOVs in the range of 1 m (as with the Siemens Quadra) to 1.5 m (as with the UI Panorama) provide superb performance with practical utility and reliable operation. However, we note that many of the benefits of TB PET can be achieved at a significantly lower cost with AFOVs in the range of 0.6 m-1 m or with a sparse detector configuration (e.g., axial gaps between rings), while still surpassing the sensitivity of conventional systems and also enabling multiorgan dynamic imaging. Allowing the user to select the most cost-effective system for their clinical or research program enables sites to complement or potentially substitute for their conventional PET-CT instruments, further enabling the growth in the number of long AFOV PET systems.

It remains to be seen whether the design of long AFOV systems will evolve beyond a multiring extension of conventional systems and their technology and whether there will be a branch to alternative designs, mentioned above, to address the challenges of high-throughput clinical use or dynamic research investigations. Potentially, other imaging modalities can be integrated with these systems to provide complementary molecular information; for example, integration of a spectral CT scanner on the one end or a high-performance brain scanner on the other end of a long AFOV PET instrument would add capabilities beyond those of current instruments. Even in their current configuration, AI/DL will continue to infiltrate the workflow of image generation and become more central to synthesizing these very large, rich data streams of molecular information to extract quantitative radiomic and kinetic biomarkers. We can definitely count on continued innovation of long AFOV PET systems to further evolve from an advanced imaging modality into a central node in systems medicine.

Conflicts of interest

J.S.K. is PI of a sponsored research agreement with Siemens Molecular Solutions.

Funding

This work was supported by the National Institutes of Health under Grants R01 CA113941 and R01 CA291948. F.M.M. acknowledges support from the Research Foundation Flanders (File number: 11P0E24N0).