Evaluation of SUVlean consistency in FDG and PSMA PET/MR with Dixon, James and Janma based lean body mass correction

Purpose To systematically evaluate the consistency of various standardized uptake value (SUV) lean body mass (LBM) normalization methods in a clinical positron emission tomography/magnetic resonance imaging (PET/MR) setting. Methods SUV of brain, liver, prostate, parotid, blood and muscle were measured in 90 18F-FDG and 28 18F-PSMA PET/MR scans and corrected for LBM using the James, Janma (short for Janmahasatian) and Dixon models. 40 dual energy X-ray absorptiometry (DXA) measurements of non-fat mass were used as the reference standard. Agreement between different methods was assessed by linear regression and Bland-Altman statistics. Results LBM fraction measured by DXA, James, Janma and Dixon approaches was 68.19±6.43%, 66.95±6.71%, 76.20±5.41% and 71.56±7.97% respectively. Compared to DXA, the Dixon approach presented the minimal bias when compared to the James and Janma models (bias: 0.76±7.35, 8.01±9.36, -3.33±8.26 respectively). SUV normalized by bodyweight (SUVbw) was positively correlated with Body Mass Index (BMI) in both FDG (liver: r=0.454, p<0.001) and PSMA studies (r=0.197,p=0.31), while SUV normalized by lean body mass (SUVlean) revealed a decreased dependency on BMI (r=0.22,0.08,0.14, p=0.04,0.46,0.18 for Dixon, James and Janma models respectively). Paired T-test showed significant difference between SUVlean of major organs measured using Dixon method vs James and Janma models. Conclusion Significant systematic variation was found among SUVlean calculated using different approaches. A consistent correction method should be applied in PET/MR serial scans.


Introduction
With the development of integrated positron emission tomography/magnetic resonance imaging (PET/MR), its comprehensive contrast mechanisms and seamless fusion of morphology and function is attracting growing clinical adoption and research exploration. Because of the reduced radiation dose of PET/MR versus PET/CT, and the increasing utilization of non-FDG (fluorodeoxyglucose) tracers, such as PSMA (Prostate Specific Membrane Antigen) and DOTATATE [1,2], serial PET/MR scans are becoming more desired for re-staging and treatment response evaluation for oncological patients. Serial PET scans require higher standards for the quantitation consistency as patients might present dramatic physiological variation in terms of body weight throughout the course of treatment [3].
In PET studies, standardized uptake value (SUV) is the most widely used semiquantitative measurement of radiotracer uptake, which is essential for diagnosis and treatment response assessment. SUV is defined as the radioactivity in a region of interest (ROI) normalized by the total injected dose and body weight of the patient [4]. Although SUV normalized by body weight (SUVbw) is the most popular metric in today's clinical setting, Zasadny et al found that it is highly dependent on patient weight and body fat content [5]. A potential cause of inconsistency is that white adipose tissue contributes to body weight but poses minimal uptake of radiotracer. SUVbw is occasionally overestimated (especially for obese subjects) and can lead to systematic bias for serial scans of patients with multiple follow-ups throughout the course of treatment [6]. Many studies have investigated improved normalization factors of SUV to allow for consistent quantitation across a wide range of body mass index (BMI) [7,8]. The most widely adopted approach is to use lean body mass (LBM) instead of body weight to offset the systematic bias caused by white adipose tissue [5]. This corrected SUV is often referred as SUVlean.
Over the past decades, various predictive models have been established to estimate LBM, taking factors such as body weight, height, sex or age into account. Some of these models have been translated into PET imaging to calculate SUVlean in the clinical setting [5,[7][8][9]. Among them, James equation [10] is the most widely used model for SUV correction and has been implemented into a large variety of commercially available PET/CT and PET/MR scanners. However, a recent study has shown that James equation might be prone to significant inaccuracy when a patient's BMI exceeds a critical value (approximately 43 for men and 37 for women) [11]. An improved model proposed by Janmahasatian et al [12] was adopted by a recent study [11] to improve the SUV consistency for patients with high BMI. However, even though model-based lean body mass estimation was derived from extensive clinical data containing a large patient cohort, there are still concerns that the predictive formula may cause substantial errors at the individual level. For instance, two patients with the same weight and height exhibit identical LBM values but may present significantly different body fat composition.
New LBM estimate approaches based on direct measurement using CT or MR in PET/CT and PET/MR [13][14][15] are believed to be more reliable than model-based LBM calculation [16]. In multi-modal PET/MR imaging, current state of the art MR based attenuation correction (MRAC) usually employs water/fat imaging using a Dixon sequence [17] and such data can be readily utilized to achieve personalized calculation of LBM. The Dixon approach was recently proposed and validated in a pilot study, where Jochimsen et al reported an initial attempt to normalize SUV with the Dixon based water/fat fraction [18]. Furthermore, without cross-validation with clinical reference standards for LBM, it remains challenging to conduct direct evaluation of the accuracy of SUVlean [19]. One approach for validating and comparing different SUV normalization models is to utilize reference standard measurements of LBM using well-established technology [16,20]. There are many ways to measure body fat and its distribution [21]. For instance, dual energy X-ray absorptiometry (DXA) utilizes different attenuation coefficient of fat and soft tissue to obtain a patient-specific fat fraction [22]. DXA has been reported to be more accurate than density-based methods and features good repeatability in which regional fat fraction can be obtained by cropping the projected 2D coronal image [23].
Although a few recent studies have reported good reproducibility of Dixon methods [24] and good agreement between Dixon and DXA measurements of body fat [15,25,26], the robustness of the PET SUV corrected by Dixon methods has not been well evaluated. In addition, because of the diverse selection of methods to calculate lean body mass for SUVlean, there is an immediate need for comparative evaluation of the consistency and limitation across these methods for SUVlean calculation in PET/MR. The purpose of the present work is to systematically evaluate the accuracy of different LBM estimation models, using DXA as a reference standard, and to investigate the consistency of various SUVlean calculation methods in a clinical setting. SUVlean measurements derived from Dixon images, as well as with James and Janma (short for Janmahasatian) LBM models were compared in two patient cohorts of 18 F-FDG and 18 F-PSMA PET/MR studies, respectively.

Patient Population
Patients (N=118) were recruited for clinical PET/MR scans from December 2018 to August 2020 at Shanghai East Hospital for suspected or known malignancies. Among them, 90 underwent 18 F-FDG PET/MR scans and 28 underwent 18 F-PSMA PET/MR scans. 40 out of 118 patients were enrolled in a same day DXA scan for body fat measurement. Patient weight ranged from 37 to 103 kg and BMI ranged from 14.53 to 32.45. Detailed information of the patients is provided in Table 1. The study protocol was reviewed and approved by institutional review board (IRB) and written informed consent was obtained from each patient.

PET/MR Image Acquisition
Whole-body PET/MR scans were performed on a hybrid PET/MR (uPMR 790, UIH, Shanghai, China), which consisted of a 3.0T MR and PET system with a transverse field of view of 60cm and axial field of view of 32cm. The PET system comprises 112 rings, each containing 700 15.5×2.76×2.76 mm 3 LYSO crystals [27]. All patients were requested to fast for at least 6 hours before the injection of the radioactive tracer. For the FDG study, patients were injected with 221±50MBq (or 0.096±0.017mCi/kg) of 18 F-FDG and rested in a quiet preparation room for about 1 hour. For the PSMA study, patients were injected with 314±73MBq of 18 F-PSMA-1007 and rested for about 2 hours. Images were acquired using the clinical PET/MR protocol at Shanghai East Hospital.

DXA Image Acquisition
DXA images were obtained using a dual-energy X-ray absorptiometry (DXA) (Lunar Prodigy, GE, Madison, USA) and body composition was analyzed using the vendor-provided software (enCORE, GE, Madison, USA). Since the PET/MR scans only covered from skull to the upper thigh, the ROI of the DXA images were adjusted to obtain the total mass and fat mass for the same axial coverage as the PET/MR images.

LBM Calculation
The total volume of water, fat, lung and air were obtained from the Dixon MRAC images by multiplying the total number of voxels by the voxel volume for each compartment. We first compared volumes of fat and water with the mass derived from DXA analysis to assess the agreement between the two measurements by linear regression. A population-based water and fat density were derived by linear fitting the volume versus measured water and fat mass from DXA. and, and compared with the direct measurement result from DXA as the reference standard.

PET Reconstruction and SUV Measurements
All PET reconstructions and image analysis were performed on the vendor-provided workstation (uPMR 790, UIH, Shanghai, China). The PET images were reconstructed using the ordered subset expectation maximization (OSEM) algorithm (FOV=600mm, iteration=2, subsets=20, Gaussian filter with FWHM=4mm, matrix size=150). Soft tissue radioactivity of major organs was measured by drawing ROIs on the workstation. Specifically, the radioactivity of the liver was obtained by placing a 3 cm diameter spherical volumetric ROI in the right hepatic lobe avoiding major vessels/lesions. The radioactivity of the blood pool was obtained by placing a 0.1 cm diameter spherical ROI in the left ventricle of the heart and the muscle SUV was obtained from the right thigh. Brain radioactivity in the FDG studies was obtained by thresholding out the whole brain using an in-house algorithm which was then subsequently validated by visual inspection. Prostate and parotid glands in the PSMA studies were detected by a thresholding tool available in the workstation software.
SUVbw was calculated by the default settings in the workstation as: ,where decay factor=exp(-0.693*wait time/radionuclide half-life).
SUV_dixon was calculated as: ,where water and fat mass were derived from the Dixon MRAC images.
SUV_james was calculated as: Finally, SUV_janma was calculated as: .

Statistical Analysis
Linear

DXA vs Dixon
The average body weight for 20 female subjects was 58.6±8.7 kg and the average body weight for 20 The slope from the linear regression was 0.78 and 1. 26 for fat and water, with r 2 =0.849 and 0.915 respectively, suggesting excellent agreement between the measurements of DXA and Dixon (Fig. 3a).
To convert MRI measured volumes to weight in the recruited patient population, lung volumes were taken into account for multi-parameter linear regression, yielding coefficients (densities) of 0.79, 1.23 and 0.20 for fat, water and lung respectively. Using the derived tissue density, the fat and water mass was determined as 17.38±5.05 kg and 33.50±6.84 kg from Dixon images. The Bland-Altman plot of fat fraction and water mass fraction are shown in Fig. 3b.

LBM Fraction
The measured LBM fraction using DXA were 63.01±3.91%, 73.37±3.71% and 68.19±6.43% for females, males and all patients. As summarized in Table 2 presented minimal bias compared to the James and Janma models.

SUV in FDG study
Linear regression between FDG SUVbw and SUVlean calculated using different approaches of LBM was plotted in Fig. 4 and the quantitative statistical results were summarized in Table 3. SUVbw was found to be significantly (p<0.05) positively correlated with BMI for all organs except for the SUVbw max value in the brain. The dependency on BMI was eliminated using Dixon, James and Janma-based SUVlean calculations as demonstrated by the reduced slope and F-test values. Paired T-test showed significant difference between SUVlean measured using Dixon approach vs James and Janma models.

SUV in PSMA study
SUVbw was found significantly (p<0.05) positively correlated with BMI for mean value within the blood pool and muscle while not statically significant within other organs measured. The dependency on BMI was eliminated using Dixon, James and Janma-based SUVlean calculations as demonstrated by the reduced slope. Paired T-test showed significant difference between SUVlean measured using Dixon methods vs James and Janma models. Linear regression between PSMA SUVbw and SUVlean is shown in Fig. 5 and quantitative statistics results are summarized in Table 4.

Discussion
Quantitation of tracer uptake in PET/MR image is essential for cancer staging and treatment response evaluation in both clinical and research settings. The pitfall of the widely used SUVbw is that it is highly dependent on patient weight and may overestimate the uptake of obese patients [5]. It has been reported that SUVbw was 70% higher in high-weight patients than in low-weight patients, and the overestimation was reduced when using other SUV normalization factors [6]. Using LBM instead of full body weight can effectively eliminate this effect and improve consistency among patients. Our findings in the FDG study were consistent to those reported by Zasadny [5] and Wahl [7,11]. Our findings in the PSMA study were also in line with previously reported results [28]. However, it is notable that, even though multiple approaches all achieved satisfactory correction for the BMI dependency of SUV, significant differences were found among Dixon, James and Janma approaches.
Using DXA as the reference standard, this study compared for the first time the accuracy of the three widely used approaches to estimate LBM. All three methods were found in good agreement with DXA, but Dixon offered the highest accuracy due to its direct measurement of body composition. James and Janma models might be prone to individual bias due to the fact that BMI might not be fully indicative of body fat content, even though they both offer reasonable population-based estimates of body fat content in the recruited cohorts from our study.
The Dixon approach offers quantitative measurement of body water/fat volume from MRI images and has gradually established itself as an alternative LBM standard [24]. Jochimsen et al [18] first proposed a method to correct SUV using water-fat signal fraction of Dixon scans in 2015. In the present work, we revised their method to be more straight-forward and easy-to-implement and validated it using a larger patient cohort and an additional tracer. Our method is different from Jochimsen's in that we utilized DXA measurements to transfer volume units into mass units, whereas they used signal intensity fraction instead.
It is notable that in previous reports [11], both body weight and BMI can be utilized as the dependent variable for SUV when evaluating the impact of obesity on SUV accuracy. We used BMI as the factor reflecting patient adiposity in our study because it is more indicative of actual fat content than body weight. For instance, patients from our cohort exhibited a wide range of body size but the most obese subject with the highest BMI was of moderate bodyweight (80kg) but very short in height (1.52m).
Therefore, weight may not be an ideal parameter to indicate the degree of obesity and we used BMI (BMI=weight/height 2 ) as an indicator of obesity instead.
It is worthwhile mentioning that even though all three LBM approaches can be utilized to correct for the BMI dependency of SUV in FDG and PSMA studies, significant variation still exists among different approaches. A potential cause is that all empirical models including James and Janma are derived from specific patient populations while Dixon is a direct measurement of body composition. In serial PET/MR scans where quantitative accuracy is crucial, a consistent SUVlean calculation approach should be adopted to correct for the change of body weight and BMI index to minimize systematic bias.
There are a few limitations in this study. Firstly, since we were using 2-point Dixon which is the basic form of Water Fat Imaging (WFI), we could only obtain the total amount of body fat by summing the total number of categorized voxels. Further work might involve the use of more advanced WFI sequences (e.g. 6-point Dixon) that can differentiate different types of adipose tissue and derive the fat content within a single voxel. Brown adipose tissue that is typically active in glucose metabolism [29,30] should not be subtracted from LBM. Secondly, although SUV dependency on BMI in the PSMA study was also positive, correlation between SUV and BMI in most organs (except for blood pool) were not statistically significant. This could be due to the limited number of patients enrolled in this study. In a recent work by Grafita et al [28], a weak but significantly positive correlation was observed between liver SUV and body weight among 121 patients who underwent 68Ga-PSMA PET/CT. The SUVlean normalized by Janma LBM was reported to have a reduced correlation with body weight. Finally, we did not include lesion uptake in this work because it is heterogenous in nature and subject to the impact of biochemistry of specific patients. The uptake of lesions depends mostly on characteristics of the tumor itself, such as tumor stage, size, degree of aggressiveness and histology type.

Conclusion
In this study we have compared LBM calculated using Dixon, James, and Janma approaches and validated their accuracy using DXA measurement. All three methods offer good estimates of LBM ratio with Dixon offering the best agreement with DXA. SUVbw was found to be positively correlated with BMI in the FDG and PSMA patient populations while SUVlean calculated using Dixon, James and Janma confirmed a decreased dependence on BMI. However, significant systematic variation was found among SUVlean calculated using different approaches, suggesting that a consistent correction method would be needed in PET/MR serial scans.