Integrated PET/MR for Resting State Functional and Metabolic Imaging in Human Brain: What is Correlated and What is Impacted

Purpose We aimed to, for the rst time, investigate the interplay of simultaneous functional MRI (fMRI) and FDG PET using a randomized self-control protocol on an integrated PET/MR. 24 healthy volunteers underwent PET/MR scan 30 to 40 minutes after the injection of FDG. A 22-minute brain scan was separated into MRI-off (without fMRI pulsing) and MRI-on modes (with fMRI pulsing) with each one lasting for 11 minutes. We calculated the voxel-wise fMRI metric (ReHo, ALFF, fALFF and DC), resting networks, relative standardized uptake value ratios (SUVr), Patlak Ki and regional cerebral metabolic rate of glucose (rCMRGlu) maps. Paired two-sample t-tests were applied to assess the statistical differences between SUVr, Ki, correlation coecients of fMRI metrics and rCMRGlu between MRI-off and MRI-on mode, respectively. mode as compared to MRI-off mode.


Introduction
PET with 18 F-uorodeoxyglucose (FDG) has been regarded as the gold-standard technique to quantify brain energy metabolism. Despite standardized uptake values (SUV) re ecting regional FDG uptake, intrinsic networks derived from FDG-PET using independent-component analysis (ICA) could re ect long-distance metabolic brain connections. Based on this data-driven technique, characteristics of the metabolic resting state networks (RSN) could be utilized to observe cognitive decline in Alzheimer's Disease and amyotrophic lateral sclerosis [1,2]. Another promising aspect of FDG-PET is to couple glucose utilization with cerebral blood ow and oxygen consumption. Blood-oxygenation-level dependent (BOLD) functional magnetic resonance imaging (fMRI) has been commonly used to non-invasively obtain oxygen metabolism and hemodynamic response related to brain activity [3]. Neurometabolic coupling between fMRI and FDG PET metrics has provided informative biomarkers for characterizing focal differences of energy demand in a healthy population and pathology processes such as alcohol exposure and disorders of consciousness [4][5][6].
Integrated PET/MR system could simultaneously obtain both fMRI and FDG PET images [7,8]. With the advantage of precisely-matched structural localization and consistent physiological states, the combined system could provide great opportunity for multiparametric analysis of interactions between brain metabolism and function [9,10]. With the addition of FDG PET, fMRI showed potential for baseline shifts in quanti able metabolism and neuronal signalling [11,12]. Also, with the fusion of fMRI, FDG PET could track dynamic task-related haemodynamic and metabolic interactions with higher temporal resolution [13]. Recently, the integration of the two modalities has proved to be cost-effective for clinical research, especially in neurodegenerative disorders. Focal alterations between FDG uptake and fMRI metrics obtained by integrated PET/MR showed potential to reveal signalling hierarchies in hippocampal-cortical circuits and default mode networks in patients with Alzheimer's Disease [14][15][16][17][18]. Altered bioenergetic coupling across gray matter and its relationship with seizure outcomes was also reported in patients with medial temporal lobe epilepsy [19]. Thus, simultaneous PET/fMRI has become a widely adopted approach for non-invasive brain metabolic researches.
In order to shorten the time of scan, some clinical research using integrated PET/MR system prefer to run MRI sequences during the FDG uptake phase (the rst 40 minutes after radiotracer administration).
However, the acoustic, thermal and electromagnetic effect imposed by MRI on human brain may result in physiological inference on quantitative FDG uptake curve. Although it's been assessed on phantoms that the stability of PET quantitation was not affected during simultaneous MRI scan, even when aggressive sequence such as fast spin echo MR sequence was performed [20][21][22], the impact of MRI sequence pulsing was found in human studies affecting brain activation including auditory, visual and motor functional cortex [23][24][25][26][27][28]. A recent study has shown that the acoustic noise produced by fast switching gradients could cause a reproducible increase ranged from 3-9% in FDG uptake, limited to the primary auditory cortex [29]. Therefore, the timing of the fMRI scan is preferred to run during the plateau phase (30 to 40 minutes after radiotracer administration) of FDG uptake curve in clinical routine. During this period, most irreversible FDG distribution is considered to be nished with only minimal amount of free FDG in blood pool. However, few studies have proofed that the impact induced by MRI scan during this "static" period could be limited to an acceptable range in human brain. Also, whether the possible effect could in uence speci c metabolic parameter or neurometabolic coupling limited to focal brain cortex, network level or whole brain level, are largely unknown. Therefore, the principal purpose of this work is to systematically evaluate the impact of simultaneous fMRI scan on quanti cation of brain metabolism using an integrated PET/MR system. We designed a randomized self-control study on healthy volunteers using clinically relevant imaging protocols. We will focus on the additional effect induced by fMRI sequence pulsing on voxel-wise energy consumption, spatial distribution of metabolic networks as well as the correlation between regional glucose metabolism and functional metrics within and across different resting state functional networks.

Subjects
All subjects were provided with written informed consent to undergo procedures approved by the Medical Research Ethics Committee of Xuanwu Hospital, Capital Medical University (Beijing, China). 24 healthy subjects (Male, 14, Female, 10) with ages of 31 to 66 years were enrolled in this study. Subjects were all right-handed and free of personal or family history of psychiatric or neurological disease, diabetes, renalfailure, claustrophobia and other MRI-related exclusion criteria. Prior to the scan, participants were directed to fast for at least 6 hours before the 18 F-FDG injection. Immediately before the scanning began, all subjects underwent a blood glucose level measurement and all measured values were below 8 mmol/L.

Data Acquisition Protocol
Simultaneous 18 F-FDG PET/MR scanning was performed on an integrated PET/MR system (uPMR 790, United-Imaging Healthcare, Shanghai) equipped with a 24-channel phase-array head coil [20]. Phantom experiments were initially performed to evaluate the hardware-related effects, the SUV and MRI signal stability of this integrated system (See Supplementary Materials). The MRI and PET list mode data were simultaneously acquired 30 to 40 minutes after the injection of 18 F-FDG as illustrated in Fig. 1.
The MRI scan started with 5 minutes of scanning for attenuation correction and 3D anatomical localization followed by a 22-minute experimental scan. Speci cally, the rst 5 minutes of MRI scanning consisted of a localizer and an ultra-short echo time MRI sequence for PET attenuation correction, and a three-dimensional T1 weighted fast-spoiled gradient echo sequence (voxel sizes = 1.0×1.0×1.0 mm3) for brain segmentation. The following 22-minute experimental scan was separated into MRI-off and MRI-on modes with each one lasting for 11 minutes. In the MRI-off mode, no MRI sequence was performed. In the MRI-on mode, a gradient-echo echo-planar pulse sequence (repetition time = 3000 milliseconds, echo time = 30 milliseconds, ip angle = 90°, number of slices = 43, voxel sizes = 3.0×3.0×3.0 mm3, matrix size = 64×64) was performed for resting-state BOLD-fMRI measurements lasting for 10 minutes. In order to fairly evaluate the impact of BOLD-fMRI on FDG-PET, all subjects were divided randomly in to one of two groups, who received either protocol A or B. For the protocol A group, the MRI-off mode was performed ahead of the resting-fMRI, while for the protocol B group, the scan order was reversed (see Fig. 1).
The PET list mode data had a 511KeV energy drift correction applied to correct the temperature induced counting loss during MRI scanning [22]. Dynamic data was reconstructed to 10 PET image frames (2 minutes per frame, matrix size = 192×192, eld of view (FOV) = 35 cm, voxel size = 1.82×1.82×2.78 mm3) using the ordered subset-expectation maximization (OSEM) algorithm (3 iterations and 20 subsets with time-of-ight (TOF) and point-spread function (PSF)). Corrections were applied for random coincidences, dead-time, scatter and attenuation. Static data was reconstructed using all list mode events obtained during the 22-minute experimental scan (matrix size = 256 × 256, FOV = 25 cm, slice thickness = 1.4 mm, OSEM 3 iterations and 20 subsets with TOF and PSF). Whole body clinical PET/MR scan was performed after the brain scan and the SUV from aorta was used as the reference to normalize blood pool signal in Patlak Ki calculation [30].

Data Processing
Data Processing for Functional Maps Post processing for PET and MRI images is summarized in Fig. 2. PET and MRI images were registered and jointly processed using SPM12 (Statistical Parametric Mapping, https://www. l.ion.ucl.ac.uk/spm/software/spm12/).

Generation of Resting-State Functional Metrics from BOLD-fMRI
The pre-processing of resting-fMRI was performed using the Data Processing and Analysis for Brain Image (DPABI) tool (http://www.rfmri.org/dpabi) [31]. After slice timing correction, the alignment correction was applied to reduce the impact of head motion. FMRI images were spatially transformed to the characteristic T1-weighted structural template. The transformed fMRI images were resampled to 3 × 3 × 3 mm3, and nuisance covariates (24 head motion parameters, cerebrospinal uid signal, white matter signal and linear trend) were regressed out. Voxel wise fMRI metric maps were generated, including regional homogeneity (ReHo), amplitude of low frequency uctuations (ALFF), fractional amplitude of low frequency uctuations (fALFF) and degree centrality (DC), using the DPARSF package inside DPABI. Eight resting-state networks (including visual, default, dorsal attention, auditory, ventral attention, control, sensorimotor and limbic network) were generated from the same corrected fMRI images using a group ICA analysis package (https://www.nitrc.org/projects/gift) and were used to mask out the regions of interest (ROI).

Calculation of Parametric Images from FDG-PET
After normalization of focal FDG uptake for body weight and injected dose, the relative standardized uptake value ratios (SUVr) with reference to cerebellum and white matter were calculated respectively. To quantify the change of SUVr during the time course of MRI-on and MRI-off mode acquisitions, Patlak Ki was derived from 2 min time frame brain PET image using a population-based input function and normalization factor from the aorta SUV of whole-body PET image. Analysis of the relative cerebral metabolic rate of glucose (rCMRGlu) at each voxel was performed [32]. To allow consistent dynamic analysis, all PET dynamic images were aligned to the rst time-point image and co-registered to the corresponding T1-weighted image. The voxel-wise Spearman correlation coe cient (R) between rCMRGlu and fMRI metrics was calculated using the python package Scipy (https://pypi.org/project/scipy/).

ICA Analysis for FDG-PET
Spatial ICA of the pre-processed PET images was performed within the MRI-off and MRI-on subgroups separately using the GIFT toolbox (http://mialab.mrn.org/software/). The optimal number of components of the principal component analysis was set to ve which was estimated using the GIFT dimensionality estimation tool. Group ICA was used to derive task-based regional activation treating MRI-on mode as a paradigm. First, the mixing matrix was estimated, which has a unique partition for each object. The component graph of each task was calculated by projecting a single task data onto the inverse of the mixing matrix partition. Task-speci c time courses and images were used to make group and inter-group inferences [33]. Each spatial map was converted to Z-values and activation maps were visualized using binary masks generated with a threshold of 1.5.

Statistical Analysis
Statistical analysis was performed using DPABI and Statistical Product and Service Solutions (IBM SPSS version 20). Paired two-sample t-tests were applied to assess the signi cance of SUV, SUVr, Ki of Patlak model between MRI-off and MRI-on mode, respectively. The voxel-wise paired t-test between whole brain SUVr in MRI-off mode and MRI-on mode was performed using DPABI. Comparisons within correlation coe cients of functional metrics (ReHo, ALFF, fALFF and DC) and rCMRGlu among the eight networks between MRI-on mode and MRI-off mode were calculated with one-way ANOVA analysis and paired twosample t-test, respectively.

MRI Impact on FDG Uptake
Phantom results have veri ed the system stability for simultaneous PET/fMRI acquisition ( Fig. S1 and Table S1). Static PET images from MRI-off and MRI-on mode produced similar SUVr(SUV WB /SUV CB )distribution maps ( Fig. 3A and 3B). From visual inspection, concordant distributions were found without focal areas of abnormal hyper-or hypo-metabolism. Voxel-wise paired t-tests between whole brain SUVr in MRI-off mode and MRI-on mode revealed no statistical difference (P 0.05), while Patlak Ki signi cantly increased when comparing MRI-on mode with MRI-off mode (P 0.001, Fig. 3C and 3D). Mean SUVr and Patlak Ki values for whole brain (WB) and different anatomical structure including grey matter (GM), white matter (WM), cerebellum (CB) and eight functional networks were calculated (Fig. 4). Mean Patlak Ki signi cantly increased when comparing MRI-on mode with MRI-off mode, while mean SUVr showed no statistical difference (P 0.05). The same measurement using white matter as a reference tissue was performed and a similar result was con rmed (Fig. S2).

MRI Impact on Metabolic Network
The most active network components derived from PET static spatial ICA were located in the auditory, default, visual and language networks (Fig. 5A and 5B). From visual inspection, there was little difference observed between the component maps calculated from PET image generated during MRI-off and MRI-on mode. Group ICA which treated fMRI scan as a stimulating paradigm revealed that the most active network components were frontal pole, superior frontal gyrus, middle temporal gyrus and occipital pole (Fig. 5C).

Impact of Simultaneous Scan on Neurometabolic Coupling
Spatial correlation maps between rCMRGlu and four fMRI metrics, as well as the results of the correlation analysis over each of the eight resting-state networks are shown in Fig. 6. Overall, high correlation coe cients were found among the four-fMRI metrics with rCMRGlu in MRI-off and MRI-on mode (mean R for ALFF, fALFF, DC and ReHo, 0.195 ± 0.260, 0.296 ± 0.180, 0.287 ± 0.164, 0.413 ± 0.145, P = 0.022). ReHo provided signi cantly higher correlation coe cients with rCMRGlu, compared to other metrics (ReHo and ALFF, P = 0.013, ReHo and fALFF, P 0.001, ReHo and DC, P = 0.002). No difference was found between other metrics (ALFF and fALFF, P = 0.144, fALFF and DC, P = 0.686, ALFF and DC, P = 0.080). The highest correlation coe cients between rCMRGlu and all fMRI metrics were found in the visual network (mean R, 0.523 ± 0.057) and default network (mean R, 0.461 ± 0.099).

Discussion
To the best of our knowledge, this work is the rst study systematically assessing the impact of simultaneous fMRI scan on FDG-PET in human brain with an integrated PET/MR system. Our protocol is a self-control study design following the recommended clinical routine, where MR sequences were performed 30 to 40 minutes after the injection of 18 F-FDG. It's commonly considered that FDG-PET data acquired during this plateau phase of uptake curve mainly represents the neuronal activity occurred during the preceding early uptake phase. Thus, SUVr for plateau phase is supposed to be steady even when fMRI scan is synchronously performed. Our results found no difference in neither mean SUVr nor voxel-wise SUVr compared between MRI-on and MRI-off period, which supported this inference. However, obvious increase of Patlak Ki was detected in MRI-on period across the whole brain, including grey matter, white matter, cerebellum and eight speci c functional networks. This unspeci c global metabolic change re ects a short-term FDG uptake elevation due to the fMRI scan, but may not in favour of a neuronal origin, whose activation should mainly locate in grey matter regions. We assume that this phenomenon may due to other physiological inference, such as temperature-dependent acceleration of metabolic rates.
Thus, with only minor amount of free FDG in blood pool available, fMRI scan in the steady state could only produce limited and unspeci c effect on the trend of FDG elevation, which does not affect the calculation of SUVr value to statistical signi cance.
In the network-wise comparison, when treating fMRI scan as a stimulation task, the effected components were located in the default, auditory, visual and language networks, which were commonly regarded as "higher-order" cognitive networks. A previous study reported that 13 meaningful RSNs could be detected from FDG-PET data acquired 10 minutes to 30 minutes post-injection. Among them, seven networks could be detected by both modalities, including default mode, left central executive, primary and secondary visual, sensorimotor, cerebellar, and auditory networks [34]. In our study, the "activated" networks induced by fMRI scan fundamentally located in these "dually" detected RSNs, which could be explained by changes of either cerebral blood ow or activity-dependent glucose consumption. As our data was acquired 30 minutes post-injection, the contribution of a blood ow signal change caused by the instant injection of FDG could be negligible. In this way, these "activations" observed in our ICA results should be regarded as comparable elevated glucose consumption, rather than increased cerebral blood ow or oxygen consumption.
In previous studies, the possible mechanism underlying neurometabolic coupling was explained by temporally synchronized cerebral blood ow and energy utilization, based on the theory that resting-state glucose and oxygen metabolism were closely linked [35]. Strong coupling was found in default and visual networks, while weak correlation was found in limbic and somatomotor network [36]. The highest correlation between rCMRGlu and fMRI metrics were achieved in ReHo, which was both detected by using integrated PET/MR or separated PET and MRI devices [3,36,37]. Other metrics, such as ALFF or DC, showed lower association with CMRGlu, maybe strongly affected by venous vasculature or other nonneuronal factors on signal amplitude [3]. These differences could also be explained by the different physiological phenomena probed by each metric. ALFF contrast is only due to single voxel signal. ReHo could be considered as a measurement of short-range FC affected by neighbouring 27 voxels, while DC measures distant voxels weighted by long-range FC in the whole brain.
Our study reports a synthesis effect of MRI scan on quantitative PET in an integrated system. Among all possible factors, acoustic MRI noise resulting from echo-planar imaging (EPI) should be regarded as the main concern. This sequence is normally accompanied by a gradient-shifting noise with sound level greater than 100dB [28]. Studies have focused on measuring how background acoustic noise in uence the hemodynamic responses in auditory cortex and made efforts to spoil the interference [24]. Reduced activation in the visual cortex was also reported, which may relate to attention modulation due to auditory-visual cross-modal neural interaction [28]. Increased activation of working-memory network [27], as well as suppressed activation in the default-mode network and sensorimotor cortex [25,26], were respectively discussed under the presence of BOLD-related noise. "Quieter" fMRI acquisition methods, such as sparse temporal sampling or interleaved silent steady state, could be applied to a less noisy background environment for BOLD-fMRI scan [23]. In addition, MR-induced RF power deposition and the resulting effects on temperature-dependent metabolic rates could also in uence FDG uptake, with maximum relative increases of 26% for uptake models based on metabolism [38]. We speculate that these above factors synergistically in uenced brain metabolism during the static phase of FDG uptake in our study.
This work was subject to several limitations. First, we adopted a blood-free approach to estimate the relative quanti cation of CMRGlu, which is more tolerable for a universal clinical routine. However, for a more precise design, absolute quanti cation of CMRGlu could be calculated by infusion of 18 F-FDG and venous blood sampling [39]. Second, methodologically, ICA and seed-based functional connectivity (sbFC) are two main approaches for statistical mapping of RSNs derived from FDG-PET. It's been discussed that the choice of ICA or sbFC could in uence the detectability of RSNs especially when sample size is limited [40]. Future studies could retest and verify our results by different data analysis methods on the basis of a larger dataset. Figure 1 Schematic diagram of the simultaneous 18F-FDG PET/MR acquisition protocol. 18F-FDG -18F-2-uro-Ddeoxy-glucose, UTE -ultra-short echo, 3D-T1W -three dimensional T1 weighted sequence. Schematic diagram of data processing protocol. The calculation of SUVr, fMRI metrics, metabolic networks and rCMRGlu maps were illustrated.  Comparison between static spatial ICA in MRI-off and MRI-on mode. Major networks from resting PET SUVr images in MRI-off and MRI-on mode were persevered (A and B). Network activation through group-ICA of combined MRI-off and MRI-on mode, treating fMRI noise as a stimulating paradigm (C).