Validation of MR-Based Attenuation Correction of a Newly Released Whole-Body Simultaneous PET/MR System

The aim of this study was to validate quantitative performance of a newly released simultaneous positron emission tomography (PET)/magnetic resonance imaging (MRI) scanner, by using MR-based attenuation correction (MRAC), both in phantom study and in patient study. PET/MRI image uniformities of a phantom under different hardware configurations were tested and compared. Thirty patients were examined with 2-deoxy-2-[18F]fluoro-D-glucose (18F-FDG) PET/computed tomography (CT) and subsequent PET/MRI. PET images from PET/MRI were corrected with MRAC (PETMR), CT-based attenuation maps (μ-maps, PETCT), and segmented CT μ-maps (PETCTSeg) derived from PET/CT. Standardized uptake values (SUVs) were compared among the 3 sets of PET in main organs (bone, liver and lung) and in 52 FDG-avid lesions, including soft-tissue lesions and bone lesions. The result showed that PET imaging uniformities of PET/MRI under different configurations were good (<8.8%). The SUV differences among the 3 sets of PET varied with organs and lesion types. In detail, the mean relative differences of SUV between PETMR and PETCT were as follows: −18.8%, bone (SUVmean); −8.0%, liver (SUVmean); −12.2%, lung (SUVmean); −18.1%, bone lesions (SUVmean); −13.3%, bone lesions (SUVmax); −8.2%, soft-tissue lesions (SUVmean); and −7.3%, soft-tissue lesions (SUVmax). The mean relative differences between PETMR and PETCTSeg were as follows: −19.0%, bone (SUVmean); −3.5%, liver (SUVmean); −3.3%, lung (SUVmean); −19.3%, bone lesions (SUVmean); −17.5%, bone lesions (SUVmax); −5.5%, soft-tissue lesions (SUVmean); and −4.4%, soft-tissue lesions (SUVmax). The differences of SUV between PETMR and PETCT were larger than those between PETMR and PETCTSeg, in both soft tissue and soft-tissue lesions (P < 0.001), but not in bone or bone lesions. In conclusion, MRAC in the newly released PET/MR system is accurate in most tissues, with SUV deviations being generally less than 10%, compared to PET/CT. In bone, however, underestimations can be substantial, which may be partially attributed to segmentation of the MR-based μ-maps.


Introduction
e integrated whole-body positron emission tomography (PET)/magnetic resonance imaging (MRI) is emerging as a potential tool in clinical practice and in medical research. To date, there have been three commercial PET/MRI systems introduced; the first system was launched by Siemens (mMR PET/MR) in 2010 featuring an integrated design for simultaneous acquisition, the second system was released by Philips (Ingenuity TF PET/MR) in 2011 using a rotating table for sequential acquisitions, and most recently, a second simultaneous PET/MR system was introduced by GE (SIGNA PET/MR) in 2016 [1][2][3].
In 2018, the United Imaging Healthcare Corporation (Shanghai, China) released a fourth system (uPMR790) to be sold commercially.
is system comprises a 3T superconducting magnet, a gradient system with a set of secondorder active shimming coil (50 mT/s, 200 T/m/s), and a 48channel radio frequency (RF) receiving system. e PET detector is installed between the gradient coil and body coil, which comprises 20 modules with a transverse field of view (FOV) of 60 cm and an axial FOV of 32 cm. Each module contains 5 × 14 blocks, and each block has 4 SiPM detector channels coupled with a 7 × 8 array of 15.5 × 2.76 × 2.76 mm 3 LYSO crystals through the proprietary design of internal light guide. e entire system comprises 112 rings, and each ring contains 700 crystal channels, making 78400 crystal channels in total. In short, the uPMR790 features 2.8 mm PET spatial resolution, the highest resolution to date, and the longest axial FOV (32 cm) of all PET/MR systems. e PET component of PET/MR systems remains challenging due to hardware integrations, and aspects of PET image reconstruction, especially attenuation correction, still remain as an issue. erefore, performance validation of this system is needed before putting it into clinical use.
Interpretation of the PET images always requires quantification of the tracer distribution, which heavily relies on attenuation correction (AC) during image reconstruction. However, the MR-based AC (MRAC) is not straightforward and quite different from computed tomography-(CT-) based AC (CTAC), as signal intensity of MRI does not reflect the electron density of the tissue, rendering a direct transformation of signal intensity to linear ACs (LACs) impossible [4][5][6]. To solve this problem, several approaches have been developed for MRAC, of which the segmentation-based methods have been commonly used and implemented into vendor-provided software [5,7]. ese methods separate the human body into several different tissue types usually using a T1-weighted 3-dimensional gradient-echo sequence with 2 echoes for fat and water separation. Afterward, predefined LACs are used for the different tissue types, allowing for computation of an attenuation map (μ-map) [5,6,8].
Several studies compared the segmentation-based MRAC against the well-established CTAC [7,[9][10][11][12][13][14][15][16]. However, conflicting results regarding underestimation or overestimation in SUVs of certain tissues remain. Furthermore, all of these comparisons were performed directly between the segmented MRAC and the continuous CTAC. erefore, the results and comparisons inevitably suffered from the impact of segmentation to the MRAC data. In addition, nonpatient objects such as MR coils within the PET FOV could also affect the accuracy of attenuation correction [9,15,17]. erefore, it is necessary to exclude the impact from these factors before validating the accuracy of the MRAC method.
is study aims at validating the performance of a simultaneous whole-body PET/MR system through two steps. In the first step, we investigated the impact of the hardware components of PET/MRI scanner (i.e., the track, patient bed, and MR coils) on the MRAC method based on the phantom study. In the second step, a volume-of-interest (VOI) based approach was conducted to validate the quantification accuracy of the MRAC-based PET in different tissues and lesions through patient-based study.

Phantom Study.
A uniform cylinder phantom (20 cm in diameter) filled with 1 mCi 68 Ge was imaged at the center of a whole-body simultaneous PET/MR system (uPMR 790, United Imaging Healthcare, Shanghai, China). Data were acquired under four different configurations ( Figure 1): (1) with the track of patient bed; (2) with the track, patient bed, and the spine coil; (3) with the track, patient bed, spine coil, and the base of the head coil; (4) with the track, patient bed, spine coil, and the whole head coil. PET data were acquired for 6 minutes under each configuration and were reconstructed by using the algorithm of time of flight (TOF) ordered subset expectation maximization (OSEM), with 20 subsets, 3 iterations, image matrix of 256 × 256, voxel size of 2.4 × 2.4 × 2.85 mm 3 , and a 3 mm Gaussian filter. e attenuation maps for rigid objects were obtained from CT scans. It is also crucial for determining the relative position between these objects and the imaging FOV. e position of the track was fixed, so its attenuation map was hardcoded into reconstruction. e positions of patient bed and coils were determined by the axial position of patient bed since coils had a fixed position on patient bed. e position of the phantom was determined with an automatic coregistration process between the build-in model and the non-attenuation-corrected PET image. e uniformity of the image was calculated based on the following approach. Five slices located at the center, ±3 cm, and ±6 cm of the image were selected from the image volume. For each slice, mean SUVs from four circles with diameters of 60 mm and a circle with a diameter of 120 mm were measured, denoted as B1, B2, B3, B4, and A1, respectively ( Figure 2). e slice uniformity is defined as e whole image uniformity is defined as the maximum of the slice uniformities from the five slices.

Ethical Statement.
e patient study was retrospective and was approved by the Institutional Review Board (IRB)/ Ethics Committee (IRB 88000039-QCN-CT5-01) in accordance with the ethical standards as laid down in the 1964 Declaration of Helsinki and its later amendments. Informed consent was obtained from all patients.

Patient Cohort.
irty patients referred for 2-deoxy-2- Inclusion criteria were as follows: (1) patients proven to have or suspected of having malignancy, with an indication for PET/CT for diagnosis, staging or restaging, follow-up, and therapy-response evaluation; (2) medical conditions of patients were stable. Exclusion criteria were as follows: (1) examinations were only on one bed position (e.g., head or abdomen); (2) obvious metal artifacts; (3) technical problems (e.g., patient movement); and (4) MR-CT registration failure. e basic information of patients is summarized in Table 1. 3.3. PET/CT Imaging. All patients fasted for at least 6 hours, and serum glucose was checked before the injection of FDG. e amount of injected radioactivity was calculated by measuring the radioactivity of the syringe before and after injection. e mean injected dose was 310.8 MBq (standard deviation (SD), 63.6). One hour after injection, PET/CT scanning was performed from the skull base to the proximal thigh in a supine position with arms over head, on the uMI 780 PET/CT scanner (United Imaging Healthcare, Shanghai, China) which had a similar PET configuration with the uPMR 790 PET/MRI scanner. Helical CT acquisition was performed without contrast enhancement using the following parameters: tube current, 274 mAs; tube voltage, 120 kV; collimation configuration, 80 × 0.5 mm; pitch, 0.516; matrix size, 512 × 512; and scanning time, 0.8 seconds per rotation. PET was acquired for 2 minutes per bed position in a three-dimensional mode, and images were reconstructed by using the OSEM algorithm. e image matrix was 256 × 256, corresponding to a 3 mm inplane pixel size with a plane thickness of 3 mm.

MR-Based Attenuation Correction.
e WFI images were preprocessed to correct for the bias field signal before submitting to segmentation. en, the corrected images were segmented into four classes: soft tissue, fat, lung, and air according to predetermined attenuation coefficients, namely, 0.096 cm − 1 , 0.080 cm − 1 , 0.032 cm − 1 , and 0 cm − 1 , in order. A deep learning-based technique was used for lung segmentation so that the segmentation could be done with arbitrary bed positions without users' input [18]. A U-Net model was used, and training data were made by arbitrarily cropping whole-body DIXON in-phase images with lung region labeled [19]. Furthermore, truncation completion was achieved with the contour of arms from PET images by segmenting non-attenuation-corrected PET images. PET images were reconstructed with a standard process provided by the vendor as follows: TOF-OSEM with 20 subsets, 3 iterations, image matrix of 256 × 256, voxel size of 2.4 × 2.4 × 2.85 mm 3 , and 3 mm post-Gaussian filter.

CT-Based Attenuation Correction.
In order to create CT μ-maps for PET reconstruction, the CT images were coregistered to the PET/MR space with a nonrigid registration algorithm using the publicly available software-Elastix [20]. e set of parameters were carefully adjusted so that the coregistration performance was optimized. We would like to point out that since nonrigid registration is underdetermined by nature, the algorithm primarily focused on good alignment in tissue boundaries between two modalities. Since CT images were acquired with arms up while MR images were acquired with arms down, the arms were excluded in the coregistration process with body-arm boundaries manually drawn for each case (Figure 3). After the coregistration process, the arms from MRAC were stitched back to CTAC to form a complete attenuation map. For each patient, two sets of CT μ-maps were obtained. One was derived directly from CT images with linear transformation. e other was segmented CT μ-maps, which was calculated as follows: for voxels of LAC <0.003 mm − 1 , they were assigned LAC � 0 mm − 1 (air); for voxels of LAC 0.003-0.009 mm − 1 , they were assigned LAC � 0.008 mm − 1 (fat); for voxels of LAC 0.009-0.011 mm − 1 , they were assigned LAC � 0.0096 mm − 1 (water); for voxels of LAC > 0.011 mm − 1 , they were assigned LAC � 0.0161 mm − 1 (bone); and for lung regions, they were assigned LAC � 0.0032 mm − 1 . e CTAC maps were used in the same manner as MRAC maps in the PET reconstruction process.
In summary, the PET dataset from PET/MRI was reconstructed using the same parameters with the MR μ-map (PET MR ), the CT μ-map (PET CT ), and the segmented CT μ-map (PET CTSeg ), as shown in Figure 4.

PET Quantification Analysis. Quantification analyses
were conducted on the vendor-provided image viewer. For measuring SUVs on the 3 sets of PET data with the same regions of interest (ROIs), the Dixon T1 images were selected and registered with all of the PET data, respectively, for providing anatomic information. ROIs in diameters of 2 cm were drawn in homogeneous area of main organs-the right lobe of the liver, the right lower lobe of the lung, and the fourth lumbar vertebrae, and SUVs were measured. Care was taken to avoid placing ROIs on large vessels, organ borders, lesions, and close to the border of the PET FOV. In addition, for analyzing the accuracy of MRAC on lesion basis, 52 FDGavid lesions from 30 patients were selected and grouped into two parts: (1) lesions located within or around bone (n = 28) and (2) lesions located in soft tissue (n = 24). VOIs with a 50% isocontour of SUV max were drawn around lesions. Mean SUVs (SUV mean ) were measured for major organs; while both SUV mean and maximum SUV (SUV max ) were measured for lesions.
is work was performed by a nuclear medicine physician with experience over 15 years.

Statistical Analysis.
Statistical analyses were performed using SPSS 20.0 (SPSS Statistics; IBM, Armonk, NY). Pearson correlation coefficient was calculated to assess the consistency between MRAC and CTAC. For each ROI or VOI, the relative difference of SUV in percent between PET CT and PET MR was defined as (SUV MR -SUV CT )/ (SUV CT ), and the relative difference of SUV between PET CTseg and PET MR was defined as (SUV MR -SUV CTSeg )/ (SUV CTSeg ). Intergroup SUV differences were analyzed using one-way ANOVA with Bonferroni post hoc correction. e significant level was set as P value less than 0.05 for all statistical analyses.

MRAC Accuracy under Different MR Hardware
Configurations. Images of cylinder phantom and calculated uniformities under different MR configurations are illustrated in Figure 1. Good image uniformities (<9%) could be obtained in all images with different MR hardware configurations. Uniformities under all MR hardware configurations met the system requirement, which was defined as <10%, demonstrating accurate MRAC for MR hardware in uPMR 790 system.

Comparisons of SUVs in Main
Organs. Mean SUVs of bone, liver, and lung all revealed high correlations (>0.95) either between PET CT and PET MR , or between PET CTSeg and PET MR , demonstrating good consistencies of SUV quantifications between MRAC and CTAC ( Table 2). e mean relative differences of SUV mean in normal bone, liver, and lung were − 18.8% (SD, 5.1%), − 8.0% (SD, 3.8%), and − 12.2% (SD, 4.6%), respectively, between PET CT and PET MR , and − 19.0% (SD, 8.3%), − 3.5% (SD, 3.6%), and − 3.3% (SD, 6.0%), respectively, between PET CTSeg and PET MR . ese differences were significant for all organs (P < 0.05). e greatest underestimation in PET MR was found in bone with a relative difference of − 18.8 ± 5.1% between PET MR and PET CT , and − 19.0 ± 8.3% between PET MR and PET CTseg . e absolute intergroup differences between PET MR and PET CT were significantly larger than those between PET MR and PET CTSeg , both in liver and in lung (P < 0.001), but not in bone (P � 0.893, Table 2).

Comparisons of SUVs in Lesions.
Of the 28 bone lesions, both mean and maximum SUVs showed high correlations (r > 0.93) between PET MR and PET CT and between PET MR and PET CTSeg , demonstrating good consistencies between MRAC and CTAC ( Table 2). e relative differences of SUV mean were − 18.1 ± 17.4% between PET MR and PET CT , and − 19.3 ± 20.4% between PET MR and PET CTSeg , while the relative differences of SUV max were − 13.3 ± 10.5% between PET MR and PET CT , and − 17.9 ± 15.1% between PET MR and PET CTSeg ; all of these differences were statistically significant (P < 0.001, Table 2). Segmentation to CTAC had no effect on SUV quantification, as intergroup differences between PET MR and PET CT were not significantly different from those between PET MR and PET CTSeg, either for SUV mean or for SUV max (P � 0.519 and 0.217, respectively; Table 2).
An even higher correlation was found in the group of twenty-four lesions in soft tissue. e correlation of SUV mean and SUV max between PET MR and PET CT (r > 0.98) and between PET MR and PET CTSeg (r > 0.97) demonstrated the excellent consistencies between MRAC and CTAC ( Table 2). e relative differences of SUV mean were measured − 8.2 ± 5.8% between PET MR and PET CT , and − 5.5 ± 6.3% between PET MR and PET CTSeg , whereas the relative differences of SUV max were − 7.3 ± 3.6% between PET MR and PET CT , and − 4.4 ± 4.0% between PET MR and PET CTSeg . All these differences were statistically significant (P < 0.001). In addition, the absolute intergroup differences between PET MR and PET CT were significantly larger than those between PET MR and PET CTSeg , both in SUV max and in SUV mean (P < 0.001, Table 2), indicating a significant impact of segmentation to CT μ-map on quantification differences between MRAC-based PET and CTAC-based PET.

Discussion
In this study, we validated the quantitative performance of a newly released simultaneous PET/MRI (uPMR 790, United Imaging Healthcare, Shanghai, China). e system presented good image uniformity in phantom measurements for a variety   of MR hardware configurations, which demonstrated the accuracy of MRAC for a variety of MR hardware options. A systematic clinical comparison showed that SUVs from PET MR , PET CT and PET CTSeg were highly correlated with each other, either from a normal tissue basis or from a lesion basis. In addition, the relative differences of SUV between MRACbased PET and CTAC-based PET were generally less than <10%. ese results indicated that the uPMR 790 system functioned with great promise to provide PET images with accurate quantification that was consistent with PET/CT. is system opens many opportunities for clinical and research utility for the large amount of Chinese population.
ere are many technical challenges in the physical integration of simultaneous PET/MR systems. One significant challenge is the hardware components, including the patient bed, the track, and the MR coils. e strong radiofrequency (RF) wave and the rapid switching gradient system may induce noise to the PET data. On the other hand, the highly sensitive MR RF may pick up noise from the digital circuit of PET. If the coil is not accounted for in the attenuation map, significant quantification bias may occur, which can be as high as 20% [21][22][23]. However, in this study, good image uniformities and quantitative accuracies (bias < 10%) were obtained in all images with different MR hardware configurations, part of which had not been discussed previously.
is result was close to the data (5.2%-8.2%) reported by Wollenweber et al. [22], who advocated that the bias introduced by the coil was less susceptible to cause clinically relevant errors. We did not study the bias caused by flexible surface coil, as it had been reported to be minimal [22,24]. e segmented MRAC method classifies the body into different types of tissue; however, the classification of bone is typically underestimated and therefore not implemented into the μ-map, whereas in CTAC, bone causes the highest attenuation and is easily accounted for. erefore, significant underestimation of PET MR in bone is expected. Eiber et al. [11], Hershah et al. [12], and Heusch et al. [13] found an underestimation of SUVs (− 30.1% on average) in PET/ MR for bones. However, as these studies were performed between PET/CT and a subsequent PET/MRI, the results might be biased from the impact of the biological clearance of the tracer. When discussing the impact of different AC methods on SUV in bone or bone lesions in the same PET dataset, Martinez-Möller et al. [7] and Eiber et al. [25] found relatively lower influences, with an average underestimation of up to − 13%. Seith et al. [9] observed a maximum SUV underestimation of − 17.3% in PET/MR in bone lesions. Our results regarding the underestimation of SUVs in normal bone (− 18.8%) and bone lesions (− 18.1% for SUV mean and − 13.3% for SUV max ) were close to these reported data. One may argue that this underestimation can lead to misinterpretation in examinations performed for assessing bone lesions. However, it should be noted that the normal uptake of FDG in normal bone is typically low, and therefore, even a small underestimation of PET MR can cause a large deviation in percentage.
For soft tissue and soft-tissue lesions, the underestimations of PET MR were relative low (− 8.0% for normal liver; − 8.2% of SUV mean and − 7.3% of SUV max for soft-tissue lesions), as compared to bone tissue. Similar results were reported in previous studies which demonstrated underestimations in SUV of up to − 11% [9,26,27].
Another area with substantial underestimation of PET MR was lung (− 12.2 ± 4.6%). Previous studies indicated that the segmentation-based MRAC might not be sufficient for reliable SUV quantification in the lung, as MRAC ignored the heterogeneity of lung intensity, probably caused by gravitation [9,10,13,28,29]. Both overestimation [9,28] and underestimation [10,28] of SUV quantification in PET MR in or adjacent to lung area had been reported. Seith et al. demonstrated an overestimation of SUV (5.7 ± 13.0%) in the anterior parts and an underestimation (− 14.0 ± 15.8%) in the posterior parts of the lungs [28]. is partially supported the result of the current study, as the right lower lobe of the lung was selected for SUV measurement in this study.
One interesting finding in this study was that the relative differences of SUV compared directly between segmented MRAC-based PET (PET MR ) and linear-CTAC-based PET (PET CT ) were larger than those between PET MR and PET CTSeg when both attenuation corrections were segmentation based. is phenomenon was substantial in soft tissue or soft-tissue lesions (P < 0.001; Table 2). is means the segmentation nature of MRAC might have partially contributed to the deviations of SUV quantifications between PET MR and PET CT . But the changes from differences of SUV between PET MR and PET CT to differences between PET MR and PET CTSeg were not significant, either in bone or in bone lesions. e underlying reason might be associated with the inability of MRAC in the identification of bone. In recent years, new techniques have been developed for MRAC with continuous LACs for bone leading to improved quantification in bone and bone lesions [30,31]. ese methods could reduce bone-related PET biases and should be used in future for whole-body MRAC.
ere are several limitations in this study. First, areas of patient arms were excluded from the analysis. Since PET/CT were acquired with arms up while PET/MRI were acquired with arms down, it is very difficult to compare MRAC and CTAC in these areas. We note that none of the lesions analyzed and none of the main organs selected were in these areas. Second, ACs of brain PETs were not evaluated because the special anatomical characteristics and the different diagnostic demands in the brain were beyond the scope of this study.

Conclusion
e simultaneous whole-body PET/MRI uPMR 790 system released by the United Imaging Healthcare Corporation (Shanghai, China) functioned with accurate attenuation corrections and SUV quantifications. e SUV deviations were generally less than 10%, compared to PET/CT. is system opens many opportunities for clinical and research utility for the large amount of Chinese population. However, in bone, SUV underestimations can be substantial, which may be partially due to the segmentation of the MR-based μ-maps.