An SPM8-based Approach for Attenuation Correction Combining Segmentation and Non-rigid Template Formation: Application to Simultaneous PET/MR Brain Imaging

Abstract

We present an approach for head MR-based attenuation correction (MR-AC) based on the Statistical Parametric Mapping (SPM8) software that combines segmentation- and atlas-based features to provide a robust technique to generate attenuation maps (µ-maps) from MR data in integrated PET/MR scanners.

Methods

Coregistered anatomical MR and CT images acquired in 15 glioblastoma subjects were used to generate the templates. The MR images from these subjects were first segmented into 6 tissue classes (gray and white matter, cerebro-spinal fluid, bone and soft tissue, and air), which were then non-rigidly coregistered using a diffeomorphic approach. A similar procedure was used to coregister the anatomical MR data for a new subject to the template. Finally, the CT-like images obtained by applying the inverse transformations were converted to linear attenuation coefficients (LACs) to be used for AC of PET data. The method was validated on sixteen new subjects with brain tumors (N=12) or mild cognitive impairment (N=4) who underwent CT and PET/MR scans. The µ-maps and corresponding reconstructed PET images were compared to those obtained using the gold standard CT-based approach and the Dixon-based method available on the Siemens Biograph mMR scanner. Relative change (RC) images were generated in each case and voxel- and region of interest (ROI)-based analyses were performed.

Results

The leave-one-out cross-validation analysis of the data from the 15 atlas-generation subjects showed small errors in brain LACs (RC=1.38%±4.52%) compared to the gold standard. Similar results (RC=1.86±4.06%) were obtained from the analysis of the atlas-validation datasets. The voxel- and ROI-based analysis of the corresponding reconstructed PET images revealed quantification errors of 3.87±5.0% and 2.74±2.28%, respectively. The Dixon-based method performed substantially worse (the mean RC values were 13.0±10.25% and 9.38±4.97%, respectively). Areas closer to skull showed the largest improvement.

Conclusion

We have presented an SPM8-based approach for deriving the head µ-map from MR data to be used for PET AC in integrated PET/MR scanners. Its implementation is straightforward and only requires the morphological data acquired with a single MR sequence. The method is very accurate and robust, combining the strengths of both segmentation- and atlas-based approaches while minimizing their drawbacks.

INTRODUCTION

Attenuation correction (AC) is one of the biggest challenges in achieving accurate positron emission tomography (PET) quantification in combined PET/magnetic resonance imaging (PET/MRI) (1, 2). The goal of MR-based AC (MR-AC) techniques is to address the challenge of converting the MR signal, related to the proton density and MR relaxation times of tissues, into linear attenuation coefficients (LACs), related to electron density, at PET energy levels (3).

The MR-AC strategies can be broadly classified into three classes: atlas- and segmentation-based approaches and techniques aimed at simultaneously estimating the emission and attenuation data. In atlas-based methods, several pairs of coregistered MR and attenuation maps (µ-maps) are combined to form a template set (4–8). Any new subject’s MR image is first coregistered to the MR template. The matched µ-map from the atlas is then warped back into the subject’s space using the inverse transformation. These methods require matching MR and transmission maps (or CT images), which are not always available and very accurate inter- and intra-subject coregistration between the training datasets. Local anatomical mismatches from the template, due to either natural or disease-related anatomical variants (e.g. tissue remodeling or surgical procedures) tend to introduce biased estimations, which are difficult to address (2). Segmentation-based approaches involve segmentation of morphological MR images into tissue classes to which approximate LACs are subsequently assigned (9–13). Local anatomical variations are usually better addressed on segmentation-based approaches, however errors are introduced due to, among other factors, the inability of these methods to classify all the different tissue classes and to the use of a handful of averaged LACs that may not represent accurately the subject-specific tissue properties (2, 13). Finally, the simultaneous emission and µ-map estimation is based on iterative methods following maximum a posteriori schemes with different correction terms (penalties) (14–17). Such approaches usually require time of flight (TOF) information and determine the attenuation sinogram up to a constant offset (2, 16, 18). Since the estimation of the LACs from the emission data is a “very ill-posed problem” (16), the iterative solution is highly dependent on both the initialization as well as the penalty terms introduced and could potentially lead to bias and inaccurate estimation (16, 18). Furthermore, in the case of tracers with very specific uptake, the PET emission data may not provide complete information to estimate accurate LACs in all the voxels (16). New transmission-based techniques have been recently redesigned to allow simultaneous emission and transmission acquisition (19). However, they involve additional radiation exposure and require TOF detector capabilities to simultaneously detect the events from the subject and the transmission source.

We present an approach for MR-AC based on the widely available Statistical Parametric Mapping 8 software (SPM8, Wellcome Trust Centre for Neuroimaging, UCL, London, UK) that allows accurate AC for brain PET studies. Our approach combines the strengths of the segmentation- and atlas-based approaches (e.g. use of accurate segmentation of tissue classes to enable precise non-rigid coregistration to form a template), which in principle could help minimize some of their specific drawbacks. The technique has been successfully tested on datasets acquired at two different institutions using integrated PET/MR scanners (Biograph mMR, Siemens Healthcare, Erlangen, Germany). The performance of the method was compared to both that of the AC method provided by the manufacturer, the Dixon-based segmented method described in (10) (Dixon-AC), as well as to that of the gold standard for PET/CT imaging, CT-based AC (CT-AC).

MATERIALS AND METHODS

Image Acquisition

Atlas Generation Data

CT and MR images of fifteen subjects diagnosed with brain tumors (glioblastoma) were retrospectively analyzed to create the template. The MR images were acquired on a 3T MAGNETOM Trio (Siemens Healthcare, Erlangen, Germany). T1-weighted 3D-MPRAGE MR images were acquired after administration of MR contrast agent (Magnevist) (see Supplementary Material for details).

All subjects underwent CT examinations within one month of their MRI scans (GE LightSpeed QX/i, Waukesha, WI). No surgical interventions were performed in the interval between the CT and MR imaging sessions. All subjects gave written informed consent and the local Institutional Review Board approved the study.

Atlas Validation Data

Sixteen CT and PET/MR datasets from subjects diagnosed with different grades of brain tumors (N=12) or mild cognitive impairment, MCI (N=4) were retrospectively included in the study and became the validation dataset. These data were collected at Copenhagen University Hospital, Denmark (N=7) and Technische Universität München, Germany (N=9). The PET and MR images were acquired simultaneously using the Biograph mMR scanner (20). All subjects gave written informed consent and the local Institutional Review Boards approved the study.

The emission data were acquired 74.5±19.9 (Mean ± SD) minutes after administration of 199.5±25.4 MBq of ¹⁸F-FDG (N=9) or ¹⁸F-FET (N=7). Data were acquired in 3D mode for 15 minutes (16.8±2.19) and were reconstructed using the 3D ordered-subsets expectation maximization (OSEM) algorithm with 3 iterations and 21 subsets, with corrections for random coincidences, variable detector sensitivity, dead time, isotope decay, scatter and photon attenuation (as described below). Images were reconstructed into a 344×344×127 matrix with voxel sizes of 2.09×2.09×2.03 mm³ (N=9) and 0.835×0.835×2.03 mm³ (N=7).

MR data were acquired simultaneously with the PET data using MPRAGE and dual echo Dixon-VIBE sequences. Low dose CT images were acquired at both sites without contrast for all subjects on the day of the PET/MR sessions using a Siemens Biograph 64 scanner (see Supplementary Material for data acquisition details).

MR and CT Image Preprocessing

All MR images were intensity normalized (MPRAGE_norm) following the method described in (21) using FreeSurfer v.5.2 (http://surfer.nmr.mgh.harvard.edu) to allow better image segmentation and therefore a better template formation.

CT images were then coregistered to the MR images using the SPM8 affine automatic coregistration. Finally, CT images were resliced (rCT) into the MR (MPRAGE) space to allow for comparisons of the µ-map generation and AC methods.

Atlas Generation

The template was formed by non-rigid coregistration of all 15 subjects from the atlas generation dataset. SPM8 was used to perform the three steps involved in the template formation: MR image segmentation into tissue classes, non-rigid coregistration of these tissue classes and, finally, warping of the CT images into the same (atlas) space (these steps are described in more detail in the Supplemental Material). The final CT template (CT_atlas) was created by averaging all the warped CT images.

Atlas Validation

In order to validate the atlas generation methodology we used the Leave One Out Cross Validation (LOOCV) approach with the fifteen datasets used for constructing the template.

Further validation of the final template (generated from all fifteen datasets) was performed using the sixteen validation subjects. In order to obtain an atlas-based CT image for a new subject, the tissue classes derived from the MPRAGE_norm image were warped into the template space and then the CT_atlas was inversely warped into the subject space. This was achieved by following the same three steps used in the “Atlas Generation” procedure: MR image segmentation (SPM8 New Segment), non-rigid coregistration of the tissue classes to the previously created Template (SPM8 DARTEL Existing Template) and finally inverse warping of the CT_atlas image into the subject space (SPM8 Create Inverse Warped). This final image (CT_iwatlas) becomes the image used for the atlas-based AC method.

Attenuation Correction

To compare our AC method to the CT-AC, the Hounsfield units of the CT_iwatlas and rCT images generated from the validation subjects were converted to LACs using the bi-linear transformation (22) and µ_atlas and µ_CT were generated. Gaussian smoothing with a 4 mm kernel was applied to these µ-maps to match the PET spatial resolution. Additionally, the µ-maps (µ_Dixon) generated using the method currently available on the mMR scanner (10) were also available for comparison. Briefly, in the case of head imaging, the images acquired with the Dixon-VIBE sequence are segmented into 3 tissue classes: air, fat and soft tissue. LACs of 0, 0.0854 and 0.1000 cm⁻¹ are then assigned to these tissue classes, respectively. Bone tissue is not segmented using this method. Attenuation correction factors (ACFs) in sinogram space were generated from all these µ-maps and used to correct the PET data for the validation subjects (named PET_atlas, PET_CT, and PET_Dixon, respectively).

Image Analysis

Voxel- and region-based analyses were performed to assess the accuracy of the µ-maps and reconstructed PET images. Only voxels included in the head mask created by performing binary and morphological operations (see Supplemental Material for details) on each of the subject’s MPRAGE_norm image head mask were used for comparisons. Bland-Altman plots, bias, variability Pearson’s correlation coefficients were calculated for all comparisons. Absolute relative changes (aRC) were defined as follows:

$$aR{C}_{I_X}(\%)=100·\frac{|I_x-I_{gs}|}{I_{gs}}$$

where X is the AC method, Dixon or atlas; I_X corresponds to either the PET images (PET_Dixon or PET_atlas) or the µ-maps (µ_Dixon or µ_atlas); and finally I_gs corresponds to the gold standard image (PET_CT or µ_CT for PET and µ-maps comparisons, respectively). Non-absolute RCs (naRC) were also calculated similarly but without taking the absolute value of the difference.

For the region-based analysis, ten GM regions of interest (ROIs) obtained from the Automatic Anatomical Labeling (AAL) template (i.e. Frontal, Insula, Cingulate, Limbic, Occipital, Parietal, Basal ganglia, Thalamus, Temporal and Cerebellum), one WM and one CSF ROI were selected to cover most of the brain (>90%). Since the frontal region is more prone to errors due to its proximity to the sinuses where bone and air segmentation is difficult, the Frontal ROI was further segmented into 28 AAL sub-regions for detailed analysis. A brain mask obtained from the subject specific GM, WM and CSF tissue probability maps was derived to calculate the voxel-based absolute differences. A brain cortex mask was additionally created for the validation dataset from the 5 outermost voxels of the brain mask (e.g. those adjacent to bone tissue) to assess the accuracy of the method in this challenging area.

Mean and maximum tumor-to-brain ratios (T2B) were calculated for all the ¹⁸F-FET patients (n=7) following the method explained in (23) (and in the Supplemental Material).

Statistical Analysis

The different methods were compared using paired Student’s t-tests. Statistical significance was considered for p < 0.05. Mean and SD in % (Mean±SD) of the aRCs and naRCs were computed across subjects. For ROI analysis, minimum and maximum percentage values (Mean±SD% [min max %]) of the average RCs across subjects were also calculated.

RESULTS

The tissue classes and template images generated using the iterative diffeomorphic image registration approach and the AAL ROIs used in this study are shown in Figure 1.

Figure 1.

Template images generated for the atlas method: (A) tissue classes at the final iteration of the diffeomorphic non-rigid coregistration method (Dartel), from left to right: GM, WM, CSF (top row) and bone, soft tissue and air (bottom row); (B) template MR (left) and CT (middle) images and AAL regions (right) used for ROI-analysis overlayed on the MR template image.

The LOOCV analysis of the 15 subjects included in the atlas generation dataset showed that the mean RC was 1.38±4.52% for all brain voxels and 0.75±1.6% for all the ROIs (range 0.16–1.78%).

Figure 2 shows µ-maps obtained with the three methods for one of the validation subjects, demonstrating excellent correspondence between µ_atlas and µ_CT, and the obvious misclassification of bone tissues in the µ_Dixon. The difference maps between the MR-based (µ_atlas and µ_Dixon) and CT-based µ-maps are shown in Supplemental Figure 1. More detailed results of the ROI-based analysis are provided in and Supplemental Figure 2.

Figure 2.

Comparison of the attenuation maps for a representative atlas-validation subject: the µ_atlas (A), µ_CT (B) and the µ_Dixon (C) are displayed in three orientations.

A comparison of the reconstructed PET images corrected with the three AC methods for one of the validation subjects is shown in Figure 3. As expected, the largest errors in the brain occurred in cortical areas, near the skull. The results of the quantitative ROI- and voxel-based analyses are summarized in Table 1. Comparable results, in terms of tendency, were observed when either absolute or non-absolute RCs were calculated (Table 1), with naRCs showing very reduced biases in all the cases. The errors in the 5 outermost voxels of the brain mask were 9.41±8.58% for PET_atlas and 28.15±13.36% for the PET_Dixon, similar to the underestimations previously reported in this area using the Dixon-based approach (24).

Figure 3.

Comparison of the PET images reconstructed with the three AC methods for one of the validation subjects: PET_atlas (A), PET_CT (B) and PET_Dixon (C). Corresponding RC images (masked to highlight only the brain area) for the PET data reconstructed with the Atlas-AC (D) and Dixon-AC (E) with respect to the CT-AC method.

Table 1.

Summary of the ROI- and voxel-based analyses change for the whole brain area.

		µatlas	µDixon	PETatlas	PETDixon
ROI-based analysis	Mean aRC [min max]	0.99±1.81 [0.19 2.76]	3.04±3.15 [1.99 4.51]	2.74±2.28 [1.54 5.42]	9.38±4.97 [4.44 16.16]
	Mean naRC [min max]	0.30±1.91 [−0.13 1.42]	0.45±3.62 [−2.24 2.19]	−1.01±2.5 [−2.25 −0.02]	−9.03±5.6 [−16.16 −1.89]
Voxel-based analysis	Mean aRC Mean naRC	1.86±4.06 0.49±4.44	4.18±6.68 −1.35±7.76	3.87±5.0 −1.18±6.21	13.0±10.25 −12.74±10.57

Bold and regular fonts are used to indicate absolute RC (aRC) and non-absolute RC (naRC), respectively. The mean RC values were obtained by averaging across all subjects. All values are given in %.

Figures 4A and B show the Bland-Altman plots for the voxel-based analyses for all the validation subjects, comparing PET_Dixon (Fig. 4A) and PET_atlas (Fig. 4B) to PET_CT. The ROI-based comparison for the 12 ROIs are presented in Figure 5. Similar results were obtained in the 28 AAL frontal sub-regions (Supplemental Table 1).

Figure 4.

Bland-Altman plots showing the voxel-based comparisons (for all the validation subjects) between: (A) PET_Dixon and PET_CT and (B) PET_atlas and PET_CT. The colorbar shows the density of voxels on the histogram grid.

Figure 5.

ROI-based analysis for all the validation subjects of the aRCs in the PET images when using the MR-based AC methods (PET_atlas and PET_Dixon) versus the CT-based approach (PET_CT).

The mean and max T2Bs and of the SUVs for the tumor and reference ROIs are given in Table 2. No significant differences were observed when comparing mean and max T2B and SUV between PET_atlas and PET_CT. Underestimation of mean and max T2B on PET_Dixon just failed to reach statistical significance when compared to either PET_CT or PET_atlas while both mean and max SUV were significantly underestimated PET_Dixon compared to either PET_CT or PET_atlas (Table 2).

Table 2.

Summary of tumor-to-brain results for the ¹⁸F-FET subset (n=7).

	PETCT	p value CT-atlas	PETatlas*	p value atlas- Dixon	PETDixon*	p value CT-Dixon
SUVmean Tumor	2.32±0.88	0.28	2.28±0.92	p<0.003	2.05±0.81	p<0.003
SUVmax Tumor	3.69±1.97	0.17	3.63±2.01	p<0.01	3.27±1.78	p<0.006
SUVmean Ref.	1.27±0.23	0.7	1.26±0.24	p<0.002	1.17±0.22	p<0.001
Mean T2B†	1.8±0.43	0.16	1.77±0.45	0.05	1.71±0.46	0.07
Max T2B†	2.82±1.14	0.11	2.78±1.16	0.05	2.7±1.16	0.06

^*Average across all ¹⁸F-FET subjects;

^†T2B = Tumor-to-brain ratio

DISCUSSION

We have proposed and implemented a head MR-AC approach for correcting the PET data acquired on integrated PET/MR scanners. In our approach, we combined features from both segmentation and atlas-based techniques to implement a robust and flexible method for brain imaging. The method is based on the widely popular SPM software and uses the diffeomorphic image registration algorithm, DARTEL, as implemented in SPM8 (25, 26). The use of six segmented tissue classes (GM, WM, CSF, soft tissue, bone and air) instead of the classic intensity-based MR images facilitates the non-rigid coregistration step: not only there is more information available to perform the registration but also the accuracy of the image segmentation enables better coregistration of the tissue classes. Since the DARTEL non-rigid coregistration method relies on these segmentations, our approach could in principle overcome local anatomical variants given appropriate segmentation and has the potential to perform more robustly than other pure atlas-based approaches in terms of both accuracy of LACs and adaptability to local anatomical features (three examples of patients from the atlas-generation dataset with large tumors close to the bone and with surgical procedures are shown in Supplemental Figures 3–5). However, further studies in even more challenging subjects are required to fully demonstrate this potential advantage.

In terms of feasibility, our approach is very easy to implement under Matlab and requires only free software (SPM and FreeSurfer) to generate a µ-map in less than 30 minutes on a standard computer.

The similar results (in terms of RC) obtained for the post-contrast atlas generation and the pre-contrast validation datasets (Table 1) suggest that the use of contrast-enhanced MR images to generate the template does not affect the outcome. The differences between the mean RCs obtained from the analysis of the AC maps and the PET reconstructed images (Table 1) are due to the 3D cumulative effect that inaccurate µ-maps have on the reconstructed images.

It is worth noting that some of the differences observed could be due to inaccuracies in the CT-derived µ-map. Although currently accepted as the gold standard, artefacts in the CT images can affect CT-AC, being one of the sources of errors in PET quantification (27). For example, artefacts commonly appear around metallic dental fillings, implants, braces, etc. (e.g. see Supplemental Figure 6). In most cases these produce streak artefacts in the CT images, while in the MR images they typically generate an ovoid-shaped signal void due to the magnetic susceptibility. While the CT streak artefacts directly affect the resulting µ-maps, they do not appear in the maps derived using the atlas-based method (Supplemental Figure 6). However further studies are required to fully understand the potential benefits and drawbacks of the MR-based approach in this context.

While SUVmean (and SUVmax for tumor ROIs) were significantly underestimated for both tumor and reference ROIs when comparing PET_Dixon to either PET_CT or PET_atlas, mean or maximum T2B also showed a consistent underestimation but just failed to reach statistical significance (Table 2). This could be due to two reasons: first, both reference and tumor ROIs were underestimated for PET_Dixon in terms of SUV, and therefore by taking their ratio, some (but not all) of the underestimation effects were compensated; and second, the small sample of subjects analyzed (n=7) especially considering the first point (a compensation of underestimation factors when taking the ratio) prevented the T2B from reaching statistical significance. Due to the fact that bone is not considered in the Dixon approach, highly (statistically) significant underestimations occur on the PET_Dixon images that could severely bias the quantification and potential clinical diagnosis.

The results presented suggest that our method provides more accurate quantification than previously proposed MR-AC methods, both in terms of the accuracy of the µ-maps and PET quantification. Segmentation-based approaches have shown differences (compared to either CT-AC or transmission-based AC) around 10% or higher for brain imaging (11, 12, 28). A comparison between a segmentation- and an atlas-based method with two different non-rigid algorithms was performed by Malone et al. in 10 control subjects (8). The authors reported that the atlas methods offered better results when compared to transmission-based AC, with mean differences smaller than 2%. Similar results to ours were also presented by Johansson et al. with a machine learning method based on a mixture of Gaussians (29, 30) and by Navalpakkam et al. with a support vector machine (31). The authors obtained accurate LACs with mean differences of PET reconstructed images around 2 to 3% in both studies compared to CT-AC PET image reconstruction. Their methods however required a multidimensional training image space including multiple MR images (T2/VIBE and dual echo UTEs) in addition to the corresponding CT, which may reduce their practical applicability. Additionally, these methods relied on similar MR image intensities for different subjects to produce accurate results, which may be very challenging across institutions. Kops et al. previously presented an atlas-method based on the SPM2 non-rigid registration algorithm (4) and also used in (8). The mean differences reported compared to transmission-based AC revealed an over-estimation of 9%. In contrast, our method uses 6 tissue classes instead of intensity images and a diffeomorphic image registration approach which was already shown to provide accurate results in the brain (26) and is more accurate than the coregistration algorithm included in SPM2 used in (4).

A limitation of our method is that the segmentation of the six different classes is less precise in the neck area, and hence the resulting µ-maps are less accurate in this region. This is because the segmentation step relies heavily on prior anatomical information (included in the SPM software) and the New Segment tool in SPM8 is primarily focused on the head, being less accurate in other areas. Fortunately, as only a small fraction of the lines of response traverse both the neck area and parts of the brain, the effect on the final PET reconstructed brain images is minimal. The method could be in principle adapted to other body regions assuming an accurate segmentation of tissue classes (e.g. bone, soft tissue, fat, air, lungs, etc.) can be performed.

CONCLUSION

We have presented an SPM8-based approach for deriving the head µ-map from anatomical MR images that could, in principle, be used in any PET/MR scanner. The method combines the strengths of both segmentation- and atlas-based methods while minimizing their drawbacks. Its implementation is straightforward and, unlike other approaches, only requires the data acquired with a single MR sequence (MPRAGE). Compared to the method currently available on the Biograph mMR scanner, we have shown that the PET image quantification is improved by 70%, on average. The improvement is the largest in areas adjacent to cortical bone (e.g. brain cortex), which are of interest in numerous neurological disorders. A package containing the necessary software to generate the µ-maps from newly acquired MR images will be provided to interested users upon request.