Towards fast whole-body PET/MR: Investigation of PET image quality versus reduced PET acquisition times

Purpose The trend towards faster acquisition protocols in whole-body positron emission tomography/magnetic resonance (PET/MR) arises the question of whether short PET data acquisition protocols in a whole-body multi-station context allow for reduced PET acquisition times while providing adequate PET image quality and accurate quantification parameters. The study goal is to investigate how reducing PET acquisition times affects PET image quality and quantification in whole-body PET/MR in patients with oncologic findings. Methods Fifty-one patients with different oncologic findings underwent a clinical whole-body 18F-Fluorodeoxyglucose PET/MR examination. PET data was reconstructed with 4, 3, 2, and 1 min/bed time intervals for each patient to simulate the effect of reduced PET acquisition times. The 4-minute PET reconstructions served as reference standard. All whole-body PET data sets were analyzed regarding image quality, lesion detectability, PET quantification and standardized uptake values. Results A total of 91 lesions were detected in the 4-minute PET reconstructions. The same number of congruent lesions was also noticed in the 3 and 2 minutes-per-bed (mpb) reconstructed images. A total of 2 lesions in 2 patients was not detected in the 1 minute PET data reconstructions due to poor image quality. Image noise in the blood pool increased from 22.2% (4 mpb) to 42.1% (1 mpb). Signal-to-noise ratio declined with shorter timeframes from 13.1 (4 mpb) to 9.3 (1 mpb). SUVmean and SUVmax showed no significant changes between 4 and 1 mpb reconstructed timeframes. Conclusions Reconstruction of PET data with different time intervals has shown that 2 minutes acquisition time per bed position instead of 4 minutes is sufficient to provide accurate lesion detection and adequate image quality in a clinical setting, despite the trends to lower image quality with shorter PET acquisition times. This provides latitude for potential reduction of PET acquisition times in fast PET/MR whole-body examinations.


Methods
Fifty-one patients with different oncologic findings underwent a clinical whole-body 18F-Fluorodeoxyglucose PET/MR examination. PET data was reconstructed with 4, 3, 2, and 1 min/bed time intervals for each patient to simulate the effect of reduced PET acquisition times. The 4-minute PET reconstructions served as reference standard. All whole-body PET data sets were analyzed regarding image quality, lesion detectability, PET quantification and standardized uptake values.

Results
A total of 91 lesions were detected in the 4-minute PET reconstructions. The same number of congruent lesions was also noticed in the 3 and 2 minutes-per-bed (mpb) reconstructed images. A total of 2 lesions in 2 patients was not detected in the 1 minute PET data reconstructions due to poor image quality. Image noise in the blood pool increased from 22.2% (4 mpb) to 42.1% (1 mpb). Signal-to-noise ratio declined with shorter timeframes from 13.1 (4 mpb) to 9.3 (1 mpb). SUV mean and SUV max showed no significant changes between 4 and 1 mpb reconstructed timeframes. a1111111111 a1111111111 a1111111111 a1111111111 a1111111111

Introduction
The successful integration of positron emission tomography (PET) and magnetic resonance Ultimately, a whole-body PET/MR exam could be as fast as the acquisition of the MR-based attenuation correction (AC) lasts without adding further diagnostic MR sequences to the protocol. This has been first demonstrated in an initial study by Eiber et al. [17], where the MRbased AC images (Dixon sequence) have also been used as anatomic correlate to the simultaneous PET images. Conclusion of this early study was that the fast protocol using MR-based AC only provides comparable results to a fast low-dose PET/CT exam [17]. Such a whole-body PET/MR protocol theoretically could be as fast as 5x 19 seconds, which is the duration for a current standard five-bed, low spatial resolution MR-based Dixon AC protocol [3] [18].
New technical developments such as continuously moving table MRI [19] that provide seamless whole-body PET/MR at different table speeds, furthermore, allow for more flexibility in whole-body PET/MR data acquisition and may allow for fast whole-body PET/MR imaging protocols when combined with a selection of fast MR imaging sequences [19].
In this context and in preparation of future fast whole-body PET/MR imaging protocols, the question arises whether short PET data acquisition protocols allow for reduced acquisition time while providing adequate PET image quality and accurate quantification parameters. The purpose of this retrospective study, thus, is a systematical investigation of the PET image quality and quantification parameters of clinical oncologic whole-body PET/MR protocols using 18F-Fluorodeoxyglucose (FDG) as radiotracer regarding the impact of reduced PET acquisition times. List-mode format PET data of a larger cohort of 51 patients with oncologic findings who underwent whole-body PET/MR exams was retrospectively reconstructed into different time frames to systematically simulate reduced PET acquisition times. The PET data were then analyzed regarding image quality and quantification parameters. Another associated implication of this study is to investigate the potential of reducing the injected tracer dose while keeping PET acquisition times constant.

Patient population
In this retrospective study, 51 patients with different oncologic findings that underwent a whole-body PET/MR examination were included. The patient population consists of 27 female and 24 male patients (mean age 45.5 ± 20.3 years; mean BMI 25.5 ± 5.6 kg/m 2 ). All patients underwent a clinical 18 F-FDG PET/MR for staging/restaging of malignant diseases (17 lymphoma, 10 gynecological carcinoma, 7 cancer of unknown primary, 6 breast cancer, 5 rectal carcinoma and 8 others). Of these, 22 patients were referred to a PET/MR examination only. The other 29 patients first underwent a clinically indicated PET/CT with tracer injection and subsequently an additional PET/MR examination without additional tracer injection. This resulted in a broad range of post injection times at the start of the PET/MR examination. The study was conducted in conformance with the Declaration of Helsinki and approved by the Ethics Commission of the Medical Faculty of the University Duisburg-Essen (study number 11-4822-BO), and all patients provided written informed consent before examination. Detailed patient information are listed in Table 1.

Image acquisition and reconstruction
All PET/MR examinations were performed on an integrated 3 Tesla whole-body PET/MR system (Biograph mMR, Siemens Healthcare GmbH, Erlangen, Germany). The PET/MR only exams started 65 ± 14 min (mean ± SD) after injection of radiotracer 18F-FDG. The PET/MR measurements subsequently to a PET/CT started 144 ± 36 min. The patients were administered an average radiotracer dosage of 218.8 ± 59.1 MBq (mean ± SD) (range 112-350 MBq). PET/MRI examinations were performed in caudocranial direction; on average 4-5 bed positions per patient were acquired. For each bed position in each patient, PET and MR data was acquired simultaneously for 4 minutes. The PET data was sampled in list-mode format and was then retrospectively reconstructed to 4, 3, 2, and 1-minute time intervals to simulate decreasing PET acquisition times. The reconstructed subsets were generated using the identical start time. Decay-, scatter-and random correction were executed identically for all reconstructed subsets. The 4-minute PET data served as reference standard in this study. All PET data reconstruction was performed with e7 tools (Siemens Molecular Imaging, Knoxville, USA) using OP-OSEM algorithm with 3 iterations, 21 subsets and 4 mm Gaussian filter. The resulting PET images have matrix dimensions of 344x344x127 with a reconstructed image resolution of 2.09x2.09x2.03 mm. In the retrospective data reconstructions, rigid hardware components of the PET/MR system were automatically attenuation corrected with e7 tools by using CT-based templates of the patient table, the radiofrequency (RF) spine array coil and the RF head-neck coil [20], as it is also the case for routine exams using the PET/MR system. Attenuation correction of the patient's soft tissue was performed by using standard MR-based AC-map of the patient. Therefore, a two-point (fat, water) Dixon-VIBE sequence generates a four-compartment model of the patient's tissues by segmenting into air, lung, fat and soft tissue for attenuation correction [18].
MRI data was acquired simultaneous to PET data. A standardized clinical whole-body PET/MR protocol includes 4-5 bed positions with a duration of 8-12 minutes per bed position mainly consists of five MR sequences. The used MR sequences and parameters per MR protocol may differ according to the individual indications of the patients. The MR protocol was implemented using following MR sequences ( Table 2).

Image analysis
All PET/MR data sets with different reconstructed time intervals for each patient (4, 3, 2, 1 min) were analyzed regarding image quality and lesion detectability, applying a 4-point scale. The 4-point image quality score (IQS), was defined as 0 = non-diagnostic, 1 = poor, 2 = moderate, and 3 = good. Additionally, the following quantitative PET imaging parameters were assessed: the standardized uptake values (SUV) SUV mean , SUV max in all detectable lesions, in liver and in the mediastinum as well as signal-to-noise (SNR) and contrast-to-noise (CNR). Image reading was performed by two experienced readers, a radiologist and a nuclear medicine specialist. Reading was performed in consensus. The PET images were analyzed starting with the 1 min/bed timeframe to 4 min/bed to avoid bias in lesion detectability. For quantitative evaluation SUV mean , SUV max and SUV SD were measured with OsiriX (Pixmeo SARL, Bernex, Switzerland) in patient's background, liver and each lesion. Therefore, a region-of-interest (ROI) with a fixed diameter of 1.5 cm was generated in the descending aorta to measure the mediastinal blood pool as background. A ROI with a diameter of 1.5 cm was placed in the liver, which was chosen as an organ with homogeneous FDG uptake, but without any findings in each patient, serving as reference standard. Two additional ROIs (diameter 1.5 cm) were placed in the gluteus maximus and the spleen. Image noise in the liver, the gluteus maximus, the spleen and the blood pool were calculated. To quantify lesion detectability, SUVs were obtained in all recorded lesions in each patient and for each reconstruction. Volume-of-interest (VOI) around lesions were delineated using fixed threshold set to 40% of SUV max in the lesion and SUVs were measured. All ROIs and VOIs were copied in identical planes and positions in each reconstructed time interval for each patient. SNR (1), CNR (2) and image noise for each time interval were calculated as described in detail by Yan et al. [21]: Correlations and significance for image quality parameters (SNR, CNR, image noise) and potential determinants (acquisition time, dose, post injection time, BMI) were calculated. Pvalues < 0.05 were considered to be statistically significant.

Image quality
The image quality analysis of different PET acquisition times is presented in Table 3. The image quality score decreases from 2.4 at 4 minutes per bed (mpb) over 2.1 (3 mpb) and 2.0 (2 mpb) to 1.8 at 1 mpb. Image noise in the blood pool increases from 22.2 ± 10.3% in the 4 min timeframe over 26.0 ± 12.2% (3 mpb) and 33.6 ± 10.6% (2 mpb) to 42.1 ± 14.9% in the 1 min timeframe. Also the calculated image noise in the spleen, the liver and the gluteus maximus increase with shorter acquisition times. The SNR also decreases with shorter time intervals

Quantitative image parameters
A total of 91 lesions were detected in the 4 minute PET data sets serving as reference standard. Thereof, 20 lesions were located in head/neck area, 39 in the thorax region, 8 in the abdomen and 24 lesions in the pelvis or upper thighs. An identical number of congruent lesions were also noticed in the 3 and 2 mpb reconstructed PET images. In the 1-minute PET data reconstructions, overall 2 lesions in 2 patients could not be detected due to poor image quality. Patient #33 has a lymph node lesion in the left mammaria flow area, which could not be detected in the 1 minute timeframe. In this case, the patient's BMI was 32 kg/m 2 , the applied radiotracer dosage was 3.1 MBq/kg and the post injection time was 148 minutes. The CNR is with 10.3 at 4 mpb significant lower than the average CNR (26.6). Patient #24 has a lymph node lesion left cervical, which also could not be detected in the 1 minute timeframe. In this second case, the patient's BMI was 22 kg/m 2 , the applied radiotracer dosage was 3.5 MBq/kg  1. Boxplots of image quality parameter against PET acquisition time. Boxplots of the image quality score (IQS), which is defined as nondiagnostic = 0, poor = 1, moderate = 2 and good = 3; the image noise (blood pool), the signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR) against PET acquisition time in minutes per bed (mpb). Note that the image quality slightly but progressively decreases in each boxplot towards lower PET acquisition times. Also the SNR and CNR decrease with lower acquisition times, while the image noise increases. https://doi.org/10.1371/journal.pone.0206573.g001 Fast PET/MR: Investigation of PET image quality and the post injection time was 114 minutes. The image noise here was with 58% in 1 mpb reconstruction significant higher than the average image noise (42%). The detected lesions were sorted into four different body regions: head/neck, thorax, abdomen and pelvis/upper legs. In Fig 2 an overview over SUV mean , SUV max and the SNR for these different body regions is shown. The relative difference in % is given for 3, 2 and 1 min per bed PET acquisition time in comparison to 4 mpb serving as the reference standard. In the head/neck region the differences in SUVs and SNR between 4 mpb and shorter reconstructed acquisition times is highest. In the thoracic region those differences in SUV mean are minimal, whereas in the abdominal region SUV max and SNR are minimal.
https://doi.org/10.1371/journal.pone.0206573.t004 The correlation coefficients between the image quality determinants PET acquisition time, applied dose, post injection time and body-mass-index (BMI) on PET image quality parameters signal-to-noise ratio (SNR), contrast-to-noise ratio (CNR) and image noise (blood pool) are given.

Discussion
Aiming towards faster PET/MR imaging protocols, in this study the effect of reduced PET acquisition times on PET image quality and quantification parameters in a clinical wholebody PET/MR setting has been systematically investigated. List-mode format PET data of 51 patients with oncologic findings who underwent 18F-FDG whole-body PET/MR exams was reconstructed into different time frames to simulate reduced PET acquisition times. The PET data were then analyzed regarding image quality and quantification parameters.
As expected, PET image quality decreases with shorter PET image acquisition times. The subjective image quality score decreased from the 4 mpb to 1 mpb time frames. However, the average image quality of 2 mpb PET image time frames was rated as moderate, reflecting the second best out of four scores. Only the average IQS from the 1-minute timeframe was defined as poor. Overall, none of the reconstructed acquisition time frames was classified as non-diagnostic. The image noise in the blood pool, which is a quantitative parameter of image quality, increases towards shorter acquisition times from 22% at 4 mpb to 42% at 1 mpb. SNR and CNR slightly decrease with shorter acquisition time, but remain consistently high across all acquisition times.
The t-test has shown that there are no significant changes in image quality with regard to image noise, SNR and CNR between 3 and 4 minutes per bed position PET acquisition time. Whereas, toward shorter PET acquisition times the image quality decreases significant. The standard PET acquisition time in the actual whole-body PET/MR protocol is 4 minutes per bed position. Recent whole-body ultra-fast MR protocols in our clinical routine are shortened up to 3 minutes per bed MR acquisition. Calculating with 3 minutes per bed as the shortest time for MR acquisition in current PET/MR protocols, 1 minute per bed PET acquisition time could be saved to standard PET acquisition time (i.e. 25% of the total imaging protocol's time is saved) without significant decrease in image quality. Further reduction of PET acquisition time is possible (up to 2 min per bed), but accompanied with significant losses in image quality. The image quality metrics above strongly correlate with the lesion detectability in PET images. Despite the decreasing image quality with shorter PET acquisition time, no significant changes in SUV mean and SUV max were observable. The image quality was sufficient to detect all congruent lesions in all 4, 3, and 2-minute reconstructions by consensus reading. Only 2 lesions could not be detected (false-negative) when reading the 1 minute PET reconstruction intervals due to poor image quality. In general, the 2-minute acquisition time permitted correct detection of all lesions and, therefore, provides identical diagnostic information as the longer acquisition times of 3 and 4 minutes in this study.
The interpretation of variance factors of some important biological image quality determinants yields that PET image quality is mainly influenced by two key factors: the PET acquisition time and the injected tracer activity. The simplified factor analysis shows the highly significant (p < 0.01) correlation between PET acquisition time and applied activity on PET image quality. Both, acquisition time and dose exhibit a positive correlation on SNR and CNR and a negative correlation on image noise. Whereas, the post injection time showed no significant correlation between any of the image quality parameters. In the context of our study, it seems to have no impact on PET image quality whether the patients had a consecutive PET/ CT examination, subsequently referred to PET/MR, or a PET/MR only examination. Another determinant exhibiting a significant correlation with PET image quality is the patients' BMI. The BMI reveals a significant negative correlation with SNR and a significant positive correlation with the image noise (p < 0.05). A higher BMI seems to reduce PET count statistics, and therefore PET image quality. Accordingly, for patients with a high BMI it could be beneficial to perform a 4 minutes standard PET acquisition.
Figs 4 and 5 provide example cases for best and worst rated image quality (Figs 4 and 5). Patient #10 in Fig 4 is a 14-year-old child with a rather low BMI of 16.2 kg/m 2 . The PET/MR measurement here started comparably early at 63 minutes post injection. In this example, CNR and SNR were measured higher than the average values determined in this study, whereby the image noise was lower than the average over all patients. The same trend of increased SNR and CNR and decreased image noise is observable in five further patients under 15 years, who all have a BMI lower than 20 kg/m 2 and a post injection time below 80 minutes. At the other end of the patient spectrum, Fig 5 provides an example of patient 5, who is 74 years old with a rather high BMI of 32.1 kg/m 2 . The PET/MR measurement here started comparably late at 148 minutes post injection. In this example, CNR and SNR were measured lower than the average values in this study, whereby the image noise was higher than the average over all patients. From these observations, we conclude that a low BMI and a short time interval between activity injection and PET/MR examination are optimal conditions for obtaining high image quality, SNR and CNR. This is to be expected when considering factors like tracer half-life time and countrate attenuation in tissue. This conclusion can be derived from the data in Table 4 and Fig 3. Besides the two key factors acquisition time and injected activity, PET image quality is influenced by the BMI. A high BMI, a long post injection time and less injected activity worsen the resulting PET image quality with regard to higher image noise and lower SNR and CNR values.
Our study includes 51 patients with a broad range of different oncologic findings. The study setup with retrospective reconstruction of PET data in list-mode format into different time-intervals allowed for a systematic and controlled evaluation of image quality and quantitative parameters while comparing to the fully sampled 4-minute PET data set as standard of reference. This allows eliminating numerous confounding factors that would have been present when multiple tracer injections and PET re-examinations would have been performed. Additionally, the study setup eliminated the influence of tracer wash-in and wash-out that is inherently present in all comparison studies that employ re-examinations with resulting different time points after tracer injection.
The broad range of clinical indications in this retrospective PET/MR study might be considered as a limitation. Due to the heterogeneous patient cohort a large variance of some important biological image quality determinants are given. Besides the PET acquisition time also the injected activity, patients' BMI and the time between injection and examination etc. might have an impact on PET image quality. Thus, the power of statistical analyses in patient data set with such variance may be inherently low. In this patient cohort the decrease of PET image quality due to shorter PET acquisition time (2 min or less) was significant. Also the correlation of PET acquisition time, dose and BMI on PET image quality was proven to be significant in this patient cohort.
The simple gating of PET list-mode data, where only the first part of the 4-minute measurements was reconstructed into shorter 3, 2, 1-minute time intervals could be considered a constraint of this study. Instead of progressively shortening the acquisition time intervals from 4-minute to 1-minute intervals as in this study, the method of retrospective randomized undersampling of list-mode would additionally account for potential patient motion or in cases where high tracer dynamics are to be expected [22] [23].
Several recent studies have investigated the possibility to reduce the number of potentially redundant MR sequences with the overall aim to shorten the overall PET/MR acquisition time while maintaining diagnostic PET/MR information [11]- [17]. Nevertheless, the PET/MR acquisition time in these fast protocols was still 4 minutes per bed position, so that simultaneous MRI can still be considered as the time limiting factor. The ultimate short whole-body PET/MR protocol is the dual use of the Dixon AC sequence also as anatomic MR imaging correlate without adding further diagnostic MR sequences. Such a fast protocol has been investigated by Eiber et al. [17]. The conclusion was, that the fast protocol using MR-based AC only provides comparable results as a low dose PET/CT [17]. Such a protocol, theoretically, could be as fast as 19 seconds per bed, the duration for a low spatial resolution MR Dixon A protocol [3] [1]. In such a scenario, PET might be the time limiting factor in PET/MR. In the present study we have investigated that 2 minutes PET acquisition per bed position is sufficient to provide accurate lesion detection and adequate image quality. Further reduction of PET acquisition time in a comparable setting may result in the loss of diagnostic information due to poor PET image quality.
Another indication of this study is that radiotracer dose may be further reduced in 18F-FDG whole-body PET/MR in patients with oncologic findings while maintaining high image quality and accurate PET quantification. Generally, PET image quality is influenced by mainly two key factors: the PET acquisition time and the injected tracer activity. The comparatively long PET/MR examination times of current protocols may be transformed into an advantage since longer PET acquisition times provide better count statistics and image quality. Increasing the PET acquisition time to match the prolonged MR examination times may allow a decrease of the applied radiotracer activity. Oehmigen et al. [24] have shown in systematic NEMA phantom experiments under controlled conditions that reducing the PET acquisition time by half provides equivalent PET image quality as doubling acquisition time with half the injected radiotracer dose [24]. According to the results of the present study, a reduction of the radiotracer dose by a factor of two seems practical when extending the PET acquisition time from 2 min to 4 min while maintaining all other PET imaging parameters and confounding factors.

Conclusion
Reconstruction of whole-body PET data with different time intervals has shown that 2 minutes instead of 4 minutes acquisition time per bed position is sufficient to provide accurate lesion detection, high image quality, SNR and CNR in a clinical PET/MR setting despite the expected trends to lower image quality with shorter PET acquisition times. These findings provide the foundation for the potential to reduce the PET acquisition time in fast 18F-FDG whole-body PET/MR imaging protocols. Beside the two key factors PET acquisition time and applied activity, PET image quality is also influenced by the patients' BMI.
Supporting information S1