Dixon sequence with superimposed model-based bone compartment provides highly accurate PET/MR attenuation correction of the brain

Abstract

Simultaneous PET/MR of the brain is a promising new technology for characterizing patients with suspected cognitive impairment or epilepsy. Unlike CT though, MR signal intensities do not provide a direct correlate to PET photon attenuation correction (AC) and inaccurate radiotracer standard uptake value (SUV) estimation could limit future PET/MR clinical applications. We tested a novel AC method that supplements standard Dixon-based tissue segmentation with a superimposed model-based bone compartment.

Methods

We directly compared SUV estimation for MR-based AC methods to reference CT AC in 16 patients undergoing same-day, single ¹⁸FDG dose PET/CT and PET/MR for suspected neurodegeneration. Three Dixon-based MR AC methods were compared to CT – standard Dixon 4-compartment segmentation alone, Dixon with a superimposed model-based bone compartment, and Dixon with a superimposed bone compartment and linear attenuation correction optimized specifically for brain tissue. The brain was segmented using a 3D T1-weighted volumetric MR sequence and SUV estimations compared to CT AC for whole-image, whole-brain and 91 FreeSurfer-based regions-of-interest.

Results

Modifying the linear AC value specifically for brain and superimposing a model-based bone compartment reduced whole-brain SUV estimation bias of Dixon-based PET/MR AC by 95% compared to reference CT AC (P < 0.05) – this resulted in a residual −0.3% whole-brain mean SUV bias. Further, brain regional analysis demonstrated only 3 frontal lobe regions with SUV estimation bias of 5% or greater (P < 0.05). These biases appeared to correlate with high individual variability in the frontal bone thickness and pneumatization.

Conclusion

Bone compartment and linear AC modifications result in a highly accurate MR AC method in subjects with suspected neurodegeneration. This prototype MR AC solution appears equivalent than other recently proposed solutions, and does not require additional MR sequences and scan time. These data also suggest exclusively model-based MR AC approaches may be adversely affected by common individual variations in skull anatomy.

INTRODUCTION

Integrated PET/MR is a new imaging technology that has many practical benefits for patients, referring physicians and radiologists. Integrated PET/MR has potential to impact future clinical and research studies (1). Unfortunately, unlike CT (or earlier rotating transmission sources), MR signal does not provide a direct, linear relationship to electron density to calculate an attenuation coefficient map (μ-map) for 511 keV photons that is required for attenuation and scatter correction in PET (2,3). Currently, attenuation correction (AC) maps in clinical PET/MR studies of the head are derived using the Dixon sequence which provides up to 4 tissue classes, i.e. air, fat, lung and soft tissue (4,5). However, the Dixon method does not include a bone compartment leading to an underestimation of SUVs in the brain compared to the reference CT-based AC in PET/CT. In the brain, inaccurate SUV estimation for various radiotracers from integrated PET/MR may limit its research potential and reduce clinical sensitivity for subtle findings.

Proposed technical improvements in AC for integrated PET/MR systems derive attenuation information either from the PET or MR data. The emission image and the μ-map can be reconstructed simultaneously (6), potentially using prior MR data providing anatomical information. However, a unique solution only exists in cases of time-of-flight PET data. MR-based AC solutions can be divided into segmentation or atlas-based methods. Segmentation approaches assign linear attenuation coefficients (LACs) for different tissue classes after segmentation of a Dixon (4) or ultrashort-echo-time image (7-9). Dixon-only based AC approaches do not account for the bone compartment, i.e. the skull. Methods using ultrashort-echo-time detect bone, but may not clearly distinguish bone from airspaces in the skull and sinuses (10). Alternatively, atlas-based methods use anatomical models to deform or supplement a μ-map derived from the MR of an individual subject. Most such approaches rely on the construction of hypothetical CT data, i.e. pseudo-CT images are predicted from MR images (11), then MR intensities are linked to CT Hounsfield units (HUs) (12) or bone information is transferred from the CT to the MR image after comparing the MR with an existing database (13).

We report the benefits of a PET/MR AC method that supplements a conventional MR Dixon sequence-derived tissue segmentation with a superimposed model-based bone compartment. This commercially available prototype was previously evaluated for whole–body PET/MR scans excluding the brain (14). To evaluate this method, we compared SUV estimation from CT, Dixon and this model-based approach using both whole-brain and regional analyses in 16 elderly subjects being evaluated for cognitive impairment that underwent serial PET/CT and PET/MR on the same day after a single ¹⁸FDG dose. Our data demonstrate that this new method significantly reduced whole and regional brain SUV estimation bias from Dixon-based MRI.

MATERIALS AND METHODS

Patient Population

The local institutional review board approved this study and informed consent was obtained from all subjects. Sixteen patients (72.1 ± 7.5 yrs, range 58 – 85 yrs, 6 female) undergoing clinical head ¹⁸FDG PET/CT for suspected cognitive impairment were recruited to undergo a same-day, repeated-measures comparison head PET/MR without additional ¹⁸FDG radiotracer administration. An Alzheimer’s disease FDG hypometabolism pattern was diagnosed for 11 subjects (69%). After completion of the study, a board-certified neuroradiologist reviewed the radiology reports, CT and MR images to extract subject-specific features that could affect the accuracy of different AC methods. This included the extent of petrous apex, sphenoid and frontal sinus pneumatization, white matter FLAIR hyperintensities (15), the amount of dental amalgam (ordinal scale), and 3 ROIs for mean CT HUs in the clivus, basal ganglia and calvarium.

Imaging Protocol

Subjects fasted for 4 hours then were given a single intravenous injection of ¹⁸FDG (5.18 MBq/kg, mean dose = 366.3 ± 11.1 MBq) after confirmation of serum glucose below 200 mg/dL. The patients rested in a quiet room prior to undergoing a standard clinical PET/CT (Biograph mCT, Siemens Healthcare GmbH, Erlangen, Germany). From this PET/CT acquisition only the CT images were used. Patients were then transported to a nearby facility for an integrated 3-T PET/MR study (Biograph mMR, software syngo MR B18P, Siemens Healthcare GmbH, Erlangen, Germany). Time from initial ¹⁸FDG dose administration to imaging was 56.3 ± 8.7 min and 156.4 ± 37.4 min for PET/CT and PET/MR, respectively. Integrated PET/MR allowed simultaneous acquisition of multiple MR sequences during the PET list-mode acquisition. A sagittal 3D magnetization-prepared rapid gradient echo (MPRAGE, TR/TI/TE = 2300, 900, 2.77 ms, 1.2 × 1.2 × 1.3-mm resolution) for anatomic co-registration. Additional multiplanar MR sequences were obtained per standard clinical protocol.

Attenuation Correction Map Generation

Dixon μ-map

The Dixon μ-map reflects the standard 4-compartment μ-map (including air, lung, fat and soft tissue with LAC values of 0, 0.0224, 0.0854 and 0.1 cm⁻¹) from the manufacturer.

CT μ-map

The CT images acquired with the Biograph mCT were registered to the Dixon μ-map so that all images could be reconstructed from the PET emission data acquired with the Biograph mMR. Rigid registration of the CT to the MR Dixon image was done with self-written registration software using mutual information as a similarity measure. Registration was confirmed by visual inspection. The CT μ-map was then cropped using the MR-based Dixon μ-map to remove objects and patient bed from the CT image, which were not present during the scan in the PET/MR and subsequently transformed from HU to linear attenuation coefficients at 511 keV (16). Voxels not covered by the CT are filled up with voxels in the Dixon μ-map to account for potential differences to avoid influences other than from differences in the μ-maps.

Bone μ-map A

The bone attenuation map is computed based on a regular 4-compartment segmentation from a Dixon sequence. Bone information is added to these μ-maps with a model-based bone prototype segmentation algorithm (Siemens Healthcare GmbH) using continuous LACs for bone. The segmentation algorithm consists of off-line (training) and on-line (runtime) stage. The off-line stage aims to construct a pre-aligned MR model image and skull mask pair. The MR model image is carefully aligned and cropped to only include the skull relevant anatomies. The skull bone masks contain bone densities as LACs in cm⁻¹ at the PET energy level of 511 keV. In addition, a set of anatomical landmarks are defined around the skull and their detectors are trained during the offline stage. Mathematically, the detector of the ith landmark is defined as C_i(F_i(p)), where F_i and C_i denote the image appearance features calculated around voxel p and a learned Adaboost classifier, respectively. The output of the detector indicates the likelihood of voxel p belonging to the landmark.

At run-time, the MR image of the model is registered with the subject MR image. The registration algorithm consists of landmark-based similarity registration and intensity-based deformable registration. In the landmark-based similarity registration, the pre-trained detectors are used to detect a set of landmarks surrounding the skull. Specifically, the ith landmark location p_i is the voxel with the maximum detector response, defined as Eq. 1.

$$p_i=\underset{p\in I}{\max }{C}_i\left(F_i\left(p\right)\right)$$ (1)

More details can be found in (17). These landmarks are used to crop the skull area from the subject MR image in the similar way as the model MR image. Afterwards the similarity transformation between the subject and the model is derived based on the locations of these landmarks using a least-square solver. Following the similarity registration, a more sophisticated deformable registration is performed to bring the model to the subject space.

The algorithm proposed in (18) was employed for deformable registration. To achieve diffeomorphic transformation from model to subject, we decompose the overall deformation into a set of small deformations, i.e., ϕ_mdl→sub = ϕ₀ ∘ ϕ₁ ∘ ⋯ ∘ ϕ_K. Each small deformation ϕ_k is iteratively calculated by Eq 2.

$${\varphi }_k=\mathrm{I}+\epsilon \frac{\partial }{\partial \varphi }S\left({\mathit{MR}}_{\mathit{sub}},{\mathit{MR}}_{\mathit{obj}}\circ \varphi \right)$$ (2)

Here, ϕ = ϕ₀ ∘ ϕ₁ ∘ ⋯ ∘ ϕ_k−1 is the deformation derived by previous iteration. I denotes the identity mapping. S(.) defines the local cross-correlation between the warped model MR and the subject MR (18).

Note, different Dixon sequence information is employed at different stages of the registration framework. Since the first registration stage is based on anatomical landmarks, we select to use fat and out-of-phase sequences, in which the landmarks exhibit more distinctive appearance characteristics. In the second deformable registration stage, we use information from in-phase Dixon sequence as the cross-correlation calculated from this sequence is more consistent across population.

The pre-aligned skull mask is brought to the subject space following the deformation ϕ_mdl→sub. The bone density information is added to the original Dixon-based μ-map at all voxels of densities higher than soft tissue after the segmentation process. The average running time of the algorithm was 2-3 minutes per case (14).

Bone μ-map B

For Bone μ-map B, the linear attenuation coefficient for soft tissue was adapted. The original value (0.1 cm⁻¹) was optimal for whole-body 4-compartment μ-maps if the density of soft tissue is averaged throughout the body. We observed brain LACs that were 2% lower averaging 0.098 cm⁻¹. Bone μ-map B is identical to Bone μ-map A except for this lowered attenuation coefficient for soft tissue.

PET Reconstruction

From the mMR PET listmode data, only the first 10 minutes for each patient was used – this reduced the chance of artifacts due to patient motion. All PET reconstructions (OP-OSEM, 3 iterations and 21 subsets) were performed offline using JSRecon and e7tools provided by Siemens, using a 344×344×127 matrix with pixel size of 2.09 mm² and slice thickness of 2.03 mm. Next to the different human μ-maps, the corresponding hardware μ-maps were used to correct for attenuation and scatter due to the head coil and patient table. A post-reconstruction smoothing with a Gaussian filter and kernel width of 2mm full width at half maximum was applied.

Data Segmentation and Analysis

For each patient, 91 ROIs were automatically segmented on the MPRAGE using FreeSurfer v5.3 (19,20, http://surfer.nmr.mgh.harvard.edu/). The 45 brain regions for each hemisphere included cerebellar white matter and cortex, thalamus, caudate nucleus, putamen, pallidum, nucleus accumbens, hippocampus, amygdala and numerous cortical regions (FreeSurfer atlas regions #X001-X003 & X005-X035, X=1,2 for left and right respectively). The 91st FreeSurfer ROI was the unpaired brainstem. This facilitated analysis to determine which brain regions experienced the largest bias due to MR AC errors. The PET reconstructions using the Dixon μ-maps were registered to the MPRAGE using FSL (FLIRT version 5.5) (21,22) to avoid any misalignment due to patient motion. The calculated transformation matrix was also applied to the PET reconstructions where the CT, Model A and Model B μ-maps were used. The 91 ROIs were transferred to the PET images and mean SUV was calculated for each region. The percentage deviation in each region for each PET reconstruction with respect to the reference PET reconstruction using the CT μ-map was calculated.

Statistical Analysis

Paired-sample Wilcoxon signed-rank tests were used to compare the SUVs derived from the 3 MR methods (Dixon, Method A, Method B) to the SUV from CT for the same brain region and patient. In addition they were also used to compare differences in mean bias between right and left cerebral hemispheres for the 3 MR-based AC methods (relative to reference CT μ-maps). A secondary analysis was performed to identify subject specific factors that may correlate with differences in CT and MR AC methods. Spearman rank correlations characterized the association between these subject-level cofactors and the within-subject difference between the MR and CT SUVs, represented as (MR SUV – CT SUV) / (CT SUV). All statistical tests were conducted at the two-sided 5% significance level using SAS 9.3 (SAS Institute, Cary, NC).

RESULTS

MR AC SUV bias estimation

Figure 1 demonstrates typical FDG surface maps from a selected subject in this study using AC maps from PET/CT, Dixon, Model A and Model B PET/MR AC methods – temporal and parietal hypometabolism consistent with underlying AD pathology can be appreciated on surface maps derived from all methods and the images would be sufficient for clinical diagnosis. Figure 2 and Table 1 offer a global summary of the magnitude and distribution of SUV estimation biases for the 3 MR AC methods compared to reference CT AC obtained on the same day. There was a wide range of SUV estimation biases for the Dixon-based MR AC method and whole brain mean SUV underestimation. Incorporating a model of the bone compartment into the Dixon-based method reduces the magnitude and spread of regional mean bias estimation biases (Model A). Altering the LAC in Model B to reflect the attenuation of brain tissue improves accuracy – i.e. whole brain SUV estimation bias was reduced by 95% compared to Dixon alone and only 5 remaining FreeSurfer regions still had SUV estimation bias of 5% or greater (87% reduction from 38 to 5). For simplicity, the following analysis and discussion emphasize SUV estimation biases that are 5% or greater in magnitude and statistically significant (P < 0.05) compared to reference CT AC – up to 5% differences might be expected for patients on different days or different PET/CT scanners (23).

Figure 1.

Comparison of FDG PET surface maps for the left hemisphere in a subject with clinical and imaging features consistent with mild Alzheimer’s dementia (top & bottom rows = lateral and medial surfaces respectively). FDG surface map using CT AC demonstrates FDG hypometabolism in the lateral temporal-parietal regions, posterior cingulate and precuneus (1st column). Using the same SUV color scale, Dixon-based AC blunts the conspicuity of these changes, but the overall pattern can be observed once the FDG surface is re-scaled by an expert user (Dixon*). The model-based μ-maps demonstrate FDG surface maps that are indistinguishable from the CT-based attenuation data using the same SUV color scale except for subtle differences in the frontal poles. Overall, all 3 MR μ-maps can be used to make the appropriate clinical diagnosis.

Figure 2.

Line-plot of the percentage FreeSurfer regions with a given mean SUV bias for 3 MR AC methods compared to reference CT AC (16 elderly subjects evaluated for dementia, 91 FreeSurfer regions per brain). The anatomic model-based MR AC methods demonstrate narrow line shapes closer to the origin, indicating both improved precision and accuracy of the SUV estimation.

Table 1.

Summary of SUV bias for 3 MR-based AC methods compared to CT AC in 16 subjects evaluated for neurodegeneration.

MRI AC Method	Dixon	Model A	Model B
Whole brain bias	−6.4%*	2.4%*	−0.3%
Whole image bias	−5.9%*	2.7%*	0.5%
FreeSurfer ROIs+	38	13	5
Lowest ROI bias	−11.99%	−1.54%	−4.29%
Highest ROI bias	+1.49%	+12.03%	+10.48%
Top 3 regions of absolute mean bias	Lateral occipital, inferior parietal & cerebellar cortex	Pars triangularis, frontal pole & rostral middle frontal cortex	Pars triangularis, frontal pole & rostral middle frontal cortex
Asymmetries#	6	4	6
Top 3 regions of asymmetry	Lateral occipital, middle temporal & postcentral cortex	Superior temporal sulcus, posterior middle frontal & gyrus pars triangularis	Superior temporal sulcus, posterior middle frontal & gyrus pars triangularis

^*Difference compared to CT-based SUV estimation is significant (P < 0.001)

⁺Freesurfer regions with mean bias differences that were statistically significant (P ≤ 0.05) and ≥ 5% in magnitude (91 total FreeSurfer regions studied)

^#Freesurfer regions with left-right mean bias differences that were statistically significant (P ≤ 0.05) and ≥ 5% in magnitude (45 FreeSurfer regions compared between right and left)

Surface-based displays of the mean SUV bias in Figure 3 demonstrate a global and relatively symmetric 5-10% underestimation of cortical SUV throughout both cerebral hemispheres and the cerebellum for Dixon. Dixon-based SUV estimation in the medial and basal portions of the frontal and temporal lobes was accurate. Adding the anatomic model to the Dixon AC method (Model A) conversely led to SUV overestimation throughout the cortex, but of lower magnitude. The largest magnitude SUV estimation bias was located within the frontal regions. A 2% reduced LAC for Model B AC reduced the bias across the cortex and frontal regions further, but there remained some frontal-lobe-specific overestimation biases (see below).

Figure 3.

Surface maps of mean FDG PET SUV bias between CT and MR-based AC methods (N = 16 subjects, scale bar = mean bias as percentage of CT SUV). The first row demonstrates that Dixon-based AC underestimated SUV in most cortical regions, but with little bias in the basal and mesial temporal and frontal lobes. An atlas based approach (Model A) reduced overall bias, but overestimated SUV in many cortical regions. Adjusting the soft tissue LAC for brain to 0.098 cm⁻¹ (Model B) reduced bias such that only the cerebellar and rostral frontal lobes demonstrated potentially clinically significant bias (defined here as > 5% SUV estimation error).

Unlike neurodegeneration studies, interpretations of ¹⁸FDG brain studies for epilepsy are more likely to depend on the recognition of subtle visual or quantitative SUV asymmetries, often located in deep temporal lobe structures not characterized by surface projections. Figure 4 shows cross-sectional axial and coronal maps of SUV estimation bias for the 3 MR-based AC methods through the medial temporal lobe deep and superficial structures, where the majority of adult epilepsy abnormalities are found (24). The Dixon and Model B approaches show little bias in the hippocampus, amygdala, entorhinal cortex and parahippocampal gyri whereas Model A overestimates SUV uptake in these regions. Note, all 3 MR attenuation methods provide relatively symmetric data (see Table 1).

Figure 4.

Estimation of mean bias for MR-based AC methods in the medial temporal lobe structures (color scale bar = mean SUV bias compared to CT). The first panel demonstrates cropped, oblique coronal blended image of FDG-PET and MPRAGE for a patient with MR-negative right medial temporal lobe epilepsy - there is subtle 8.9% asymmetric decrease in right hippocampal SUV compared to the contralateral side (arrow). Corresponding coronal images of the mean bias maps for all 16 subjects are shown for Dixon, Model A and Model B MR AC methods. The Dixon and Model B AC maps demonstrate no clinically significant bias in the medial temporal lobe. Also note little asymmetry in SUV estimation error for all 3 methods.

Individual subject factors that correlate with MR AC error

Several individual anatomic features correlated with SUV estimations based on MR AC methods. The CT Hounsfield units in the basal ganglia negatively correlated with whole-brain mean bias for all 3 MR-based methods (e.g. for Dixon method; R = −0.69, P = 0.003). Table 2 shows the impact of frontal and sphenoid sinus pneumatization on the 3 FreeSurfer regions where Model B SUV estimation had biases compared to CT. As frontal sinus pneumatization increased amongst the 16 subjects, Model B SUV estimation error for the rostral middle frontal cortex actually decreased. Conversely, when sphenoid sinus pneumatization increased, Model B overestimations in the 3 regions increased. The frontal pole also correlated with CT Hounsfield units for the clivus (R = −0.57, P = 0.022), a potential surrogate marker for overall skull base mineralization. Otherwise, no significant correlations were detected between the 3 regions of SUV estimation error for Model B and the various other factors described in the methods (P > 0.05).

Table 2.

Sinus pneumatization correlated with mean SUV bias for Model B PET/MR AC method compared to reference CT (16 subjects, only left-sided data shown for simplicity).

FreeSurfer region	Mean SUV bias	Sphenoid sinus pneumatization	Frontal sinus pneumatization
Frontal pole	+10.5%	R = +0.56 (P = 0.024)	R = −0.36 (P = 0.175)
Rostral middle frontal	+6.1%	R = +0.46 (P = 0.073)	R = −0.55 (P = 0.027)
Pars triangularis	+7.0%	R = +0.51 (P = 0.044)	R = −0.39 (P = 0.137)

Many other FreeSurfer regions displayed correlations between Dixon SUV biases and underlying individual anatomic features that are beyond the scope of this study. Additional subject factors that were characterized (age, dental amalgam, Alzheimer’s dementia diagnosis, PET/CT or PET/MR scans time) did not have significant correlations to the mean bias for any of the MR AC methods.

DISCUSSION

This study demonstrates the benefits of modifying Dixon-based μ-maps with a model-based bone compartment superimposed using common anatomic landmarks. The model-based approach reduced whole-brain SUV estimation bias present in Dixon-only MR AC methods by 95% with residual mean SUV bias compared to reference CT AC of −0.3% (Table 1). This result remained valid for nearly all individual FreeSurfer-parcellated brain regions with only 5 of 91 FreeSurfer regions demonstrating statistically significant SUV estimation bias of 5% or greater (an 87% reduction compared to the Dixon-only method)(Figure 2 and Table 1). There were few significant SUV estimation bias asymmetries using the model-based MR AC (Table 2), a useful feature for clinical interpretation of ¹⁸FDG brain studies. The bone compartment model-based approach relies on a short Dixon sequence (19-sec acquisition) without requiring additional MR sequences. The Model B AC maps can be generated in 2-3 minutes and applied retrospectively to pre-existing data. While this study evaluated elderly subjects, the advantages of anatomy-based MR AC methods should be applicable to other patient populations common to PET studies, such as epilepsy (Figure 4).

Previous reports tried to improve MR-based AC for integrated PET/MR studies with different atlas-based approaches. A combination of local pattern recognition and atlas registration to 3 subjects resulted in a residual mean SUV bias of 3.2% ± 2.5% in 12 ROIs compared to reference CT AC (11). If PET/MR AC is based on warping individual subject MR data to a population-based atlas of coregistered CT and MR data to generate a pseudo-CT (13), the voxel-based absolute SUV estimation bias for simulated cases was 2.9% ± 0.9% and approximately 5% for a real patient case compared to CT. The Model B approach described here generated a slightly lower bias of 4.0% ± 1.5% in 16 individual subjects when using similar bias calculation methods (i.e. computed for the whole brain as segmented by FreeSurfer). Izquierdo-Garcia and colleagues used statistical parametric mapping to coregister subject PET/MR data to an anatomic template (25). Voxel-based absolute error compared to CT with this method (3.9% ± 5.0%) is equivalent to absolute whole brain bias error with Model B.

The accuracy of any anatomic-model-based MR AC regional SUV estimation may be affected by common individual-specific variations in innate skull or brain anatomy, or postsurgical changes to the skull base and calvarium. To characterize the impact of anatomic variation on model-based bone compartment modification of the Dixon MR AC method, we characterized the impact of brain, skull base and calvarial features that are known to vary amongst individuals without history of prior surgery. The largest region-specific SUV estimation biases with the Model B method were in the frontal poles and rostral middle frontal gyri similar to a previous atlas-based approach (25). We hypothesized this reflected individual variation in frontal sinus pneumatization, but a negative correlation was only present for the rostral middle frontal region (such that increasing pneumatization decreased Model B SUV overestimation bias). Conversely, Model B SUV estimation bias both for this region and the frontal poles positively correlated with sphenoid sinus pneumatization (Table 2). In a post hoc analysis, we then ranked the amount of SUV estimation bias between Model B and reference CT for the frontal poles in all 16 subjects. Visual analysis of the μ-maps demonstrated discordance between the superimposed bone compartment model and CT measured thickness of the frontal calvarium in the subjects with the largest frontal pole SUV estimation error for Model B (Figure 5). Frontal bone thickness and pneumatization are highly variable in individual subjects – this may limit pure atlas/model-based approaches to PET/MR AC.

Figure 5.

Visual comparison between CT μ-maps (A-B) and Model B μ-maps (C-D) for two individual subjects selected with high (left) and low (right) SUV bias in the rostral middle frontal FreeSurfer region. The first subject (column 1) had 8.0% mean bias (yellow region superimposed on axial MPRAGE, panel E) whereas the second subject (column 2) had 3.6% mean bias in these same bilateral frontal regions (blue region in panel F). Comparison of the μ-maps for the first subject demonstrated that MR Model B overestimated the frontal calvarium thickness (arrow) whereas the Model B μ-map (D) estimated frontal calvarial thickness more accurately for the second subject.

Several additional subject-specific features correlated with SUV estimations for all 3 MR AC methods, although not all were associated with a SUV estimation bias compared to CT. As measured CT Hounsfield units in the basal ganglia increase, whole-brain SUV underestimation for all 3 MR attenuation correction methods increases (R = −0.69 or lower, P ≤ 0.003). This requires independent verification in a larger dataset, but suggests that the optimal LAC for brain parenchyma may depend on the underlying tissue health. This result and others suggest clinical investigation for subtle SUV differences should account for limitations of the anatomic model for specific regions that vary amongst individual subjects. Further, future MR AC methods may need to derive data directly from individual patients (like UTE) and cannot solely rely on atlas-based approaches to further improve SUV estimation accuracy.

CONCLUSION

A Dixon-based MR AC with the addition of a model-based bone compartment and a 2% reduction in soft tissue LAC improved whole-brain SUV estimation accuracy by 95%. This approach gave similar or better SUV estimation accuracy improvement compared to other approaches (13,25), but represents a commercially-available prototype that does not require additional MR sequences. Besides cognitive impairment patients, this new MR AC method should increase diagnostic accuracy for other clinical groups studied with ¹⁸FDG PET (e.g. epilepsy). Residual SUV overestimation biases in the polar and lateral frontal lobe regions appear to reflect individual subject discordance between the bone compartment model and frontal calvarium thickness (not bone density or pneumatization). This suggests a model-based MR AC approach may always produce some regional biases unless modified by same-day, direct MR data that characterizes individual variation in skull anatomy well.