Introduction
Recent clinical studies have shown the advantages of Positron Emission Tomography/Magnetic Resonance (PET/MR) imaging in neuro-oncology [
1], epilepsy [
2] and neurodegenerative diseases such as Alzheimer’s disease (AD) [
3]. Correcting for photon attenuation is essential to accurately quantify the radionuclide uptake, especially in neuroimaging where the skull attenuation coefficients are high. In the absence of a transmission source, Computed Tomography (CT) image or time-of-flight PET, the attenuation information can only be derived from MRI images. However, MRI image intensities do not reflect the electron density, which prevents a direct estimation of the attenuation coefficients. With MRI images, a specific challenge is to differentiate between bone and air as they often both have low intensity. A lack of accuracy in the bone delineation has been shown to lead to a strong spatial bias of the PET activity [
4].
MRI-based attenuation correction methods can be classified in two main categories: segmentation and registration-based approaches [
5]. The segmentation-based strategy consists of assigning uniform linear attenuation coefficients to tissue classes obtained by segmenting a T1-weighted MRI image [
6], or images derived from Dixon [
7] and/or Ultrashort-Echo-Time (UTE) sequences [
8‐
12]. In the brain, the use of Dixon sequences is limited to water-fat separation while UTE sequences allow cortical bone segmentation. This class of methods can produce inaccuracies in areas such as the sinuses, where differentiation between bone and air is required, and the resulting attenuation maps do not reflect the range of attenuation values present in the body.
In registration-based methods, an attenuation map (
μ-map) template, derived from pre-acquired CT or transmission images, is deformed to match the patient’s anatomy. In the method from [
13], the CT template is directly registered to the patient MRI image while in other methods [
14‐
17], an MRI image is associated with the
μ-map template and the mapping is performed between the template and patient MRI images. Used on its own, image registration proves to be insufficient to generate accurate
μ-maps but has shown promising results when associated with a Gaussian process [
15] or a voxel-wise weighting scheme [
16,
17].
Combining segmentation and registration, Izquierdo-Garcia et al. [
18] propose to generate attenuation maps using the Statistical Parametric Mapping (SPM) software. The patient’s MRI image is first segmented into 6 tissue classes and then non-rigidly registered to a segmented MRI template. The CT template, aligned with the MRI template, is finally mapped into the patient’s space by applying the inverse transformation.
In this paper, we validate a CT and attenuation map synthesis algorithm based on a multi-atlas information propagation scheme developed by Burgos et al. [
16,
17] and compare it to a state-of-the-art method [
18] based on a package widely used in the neuro-imaging community (SPM). In order to mitigate some limitations that we observed, we also propose a better estimation of the local image similarity and an extension of the method to multi-contrast MRI data. The method was validated with 22 subjects with various dementia syndromes, who had MR imaging and two PET/CT scans with different tracers:
18F-FDG and
18F-florbetapir. The validation was done jointly using two PET tracers to confirm the independence of the method to the radiopharmaceutical administered. We assessed the accuracy of the CT synthesis by comparing to the original CT images. The PET reconstruction accuracy was assessed, for both tracers, by comparing the reference PET images, corrected for attenuation using the reference CT-based
μ-map, to the PET images corrected using the proposed method. We analysed results both across the full brain but also in regions of particular interest, e.g. when studying dementia.
Materials and methods
PET/CT and MRI data acquisition
Twenty-two sets of 18F-FDG PET, 18F-florbetapir PET, CT and MRI brain images were used in order to validate our attenuation correction method. The 22 individuals (5 with posterior cortical atrophy, 5 with semantic dementia, 4 with progressive nonfluent aphasia, 3 with logopenic progressive aphasia and 5 healthy controls) each attended for three imaging sessions on three consecutive days. At the first visit, MR imaging was acquired on a 3T Siemens Magnetom Trio scanner (Siemens Healthcare, Erlangen, Germany) and includes a T1-weighted magnetisation-prepared rapid gradient-echo (3.0 T; acquisition time 9 min 23 s; TE/TR/TI, 2.9 ms/2200 ms/900 ms; flip angle 10°; voxel size 1.1 × 1.1 × 1.1 mm3) and a T2-weighted (3.0 T; acquisition time 4 min 43 s; TE/TR, 401 ms/3200 ms; flip angle 120°; voxel size 1.1 × 1.1 × 1.1 mm3) volumetric scans. For the second and third visits, imaging was performed on a GE Discovery ST PET/CT scanner (GE Healthcare systems, Waukesha, WI), providing CT (voxel size 0.59 × 0.59 × 2.5 mm3, 120 kVp, 300 mA) and PET (voxel size 1.95 × 1.95 × 3.27 mm3) images. For visit two, images were acquired for 10 minutes, 50 minutes after injection of 200 MBq 18F-florbetapir; for visit three, images were acquired for 20 minutes, 30 minutes after injection of 185 MBq 18F-FDG. The local ethics committee approved the study and all subjects gave written, informed consent.
Because of the separate PET and MRI acquisitions, sequences used for MR-based attenuation correction, such as Dixon or UTE sequences, were not acquired. We refer the reader to [
17] for a comparison between the CT synthesis and UTE-based methods.
Validation on 18F-FDG and 18F-florbetapir PET tracers
The aim of this paper is to validate the MR-based attenuation correction method proposed by [
17] for
18F-FDG and
18F-florbetapir PET tracers. To do so, pseudo CT (pCT) images, synthesised from MRI images as explained in section “
CT synthesis”, were used during the reconstruction of PET images to correct for attenuation.
A common practice in the neuroimaging community is to normalise PET images using a reference region [
19,
20]. For the FDG PETs, the mean uptake value in the pons [
19] was used to normalise the PET images of each subject, thus allowing for a comparable range of values. For the florbetapir PETs, the mean value in the cerebellar grey matter [
20] was used. These reference regions were extracted from parcellated T1-weighted MRI images. The parcellations were obtained from a multi-label fusion algorithm, as implemented in NiftySeg [
21].
Once the PET images were normalised, the analysis consisted in providing quantitative regional assessment of the mean absolute error and the mean voxel bias between the PET images corrected with the method proposed and the PET images corrected with the attenuations maps derived from the reference CT images. The full brain region, posterior cingulate gyrus, angular gyrus, superior frontal gyrus, fusiform gyrus and anterior cingulate gyrus, regions relevant to dementia pathologies, were obtained from the parcellated T1-weighted MRI images and propagated to the PET images.
CT synthesis
The CT synthesis method developed in [
17] relies on a pre-acquired set of aligned MRI/CT image pairs from multiple subjects forming an MRI-CT database. To generate the CT from a target MRI image, each MRI image from the database are deformed to the target MRI image using affine followed by non-rigid registration [
22]. The CT images in the database are then mapped using the same transformation to the target MRI image. A local image similarity measure [
23] between the target MRI and the set of registered MRIs from the database is used as a surrogate of the underlying morphological similarity, under the assumption that, if two MRIs are similar at a certain spatial location, the two CTs will also be similar at this location. To generate the pseudo CT, the set of registered CTs is fused using a voxel-wise weighting scheme [
21,
24]. Finally, the CT values, expressed in HU, are converted to linear attenuation coefficients, in cm
−1, using a piecewise linear transformation [
25]. The resulting attenuation maps are smoothed using a Gaussian filter with a kernel standard deviation of 2 voxels (1.172 × 1.172 × 2.5 mm
3) to approximate the PET’s point spread function (PSF), and resampled to the PET’s discretisation grid.
While providing an accurate
μ-map synthesis, we found through experimentations (see section “
Results”) that the method described in [
17] had a few limitations related to the field-of-view (FOV) and the lack of complementary information, resulting in localised reconstruction inaccuracies close to critical areas used for standard uptake value ratio (SUVR) normalisation. Thus, rather than only validating the method in [
17] for two tracers, we developed novel methodologies to address these limitations.
Convolution-based ROI-LNCC
The FOV of the MRI images usually contains the head and neck of the subject, while in the CT FOV only the head is visible, which can lead to mismatching information when images from the two modalities are aligned. A similar mismatch can occur with inter-subject mapping. If the mismatch between FOVs is not taken into account, the inter-subject mapping and resampling processes introduce areas where no information is available, which can lead to severe underestimation of the
μ-map. Those areas have to be accounted for when the similarity measure is computed and during the intensity fusion process. We extend the convolution-based local normalised correlation coefficient (LNCC) method by Cachier et al.[
23] to irregular regions-of-interest (ROI) (see Appendix
A). Thanks to this process, LNCC values are only valid within the bounds of the FOV. The ROI-LNCC at each voxel is then ranked across all atlas images and the ranks are converted to weights by applying an exponential decay function. These weights are used in a spatially varying weighted averaging to reconstruct the target CT [
17].
Exploiting MRI multiple contrasts
The algorithm developed by [
17] relies on the ability to accurately map T1 brain images from different subjects, a process that can be challenging in low-contrast areas such as the sinuses and the bone/dura/cerebrospinal fluid (CSF) boundary. As T1-weighted and T2-weighted MRI sequences provide complementary information to describe the underlying subject’s anatomy, we propose to combine information from both sequences at the registration and image similarity stages.
To do so, T1 and T2 images are affinely aligned to form a T1-T2 pair. The inter-subject coordinate mapping is obtained using a symmetric global registration followed by a cubic B-spline parametrised non-rigid registration, using multivariate normalised mutual information, as implemented in NiftyReg [
22]. All the CTs in the database are then mapped to the target subject using the transformation that maps the subject’s corresponding MRI pair in the database to the target MRI. The multivariate ROI-LNCC used for the local atlas ranking procedure is here defined as the sum of the ROI-LNCC of each channel (Appendix
1B).
SPM-based approach
For comparison purposes, we implemented the approach presented by Izquierdo-Garcia et al. [
18]. The initial step consists of creating an MRI and a CT template. To do so, the T1-weighted MRI images from the MRI-CT database are first segmented into 6 tissue classes (grey matter, white matter, CSF, bone, soft tissue, and air) using SPM12 [
26]. The 22 segmented images are then non-rigidly co-registered using Dartel [
27] to form the MRI template. The same transformations are applied to the CT images from the MRI-CT database, and the CT template is created by averaging the 22 co-registered CT images. To generate a pseudo CT, the patient’s MRI image is segmented into 6 tissue classes and non-rigidly registered to the MRI template. The associated CT template is finally mapped into the patient’s space by applying the inverse transformation.
PET reconstruction
We used the PET images provided by the PET/CT scanner as input for a simulation technique. To evaluate the effect of different
μ-maps on the PET images, we followed a projection/reconstruction technique described in [
17]. The original PET image and the reference CT-based
μ-map were projected to obtain simulated sinograms. The scatter sinogram was estimated from the emission data and the attenuation maps using a single scatter simulation algorithm [
28]. Projection data were rescaled to account for attenuation and the estimated scatter sinogram was added to produce a non-corrected sinogram, similar to the data acquired by the PET/CT scanner. The non-corrected PET sinogram was then reconstructed with both scatter estimation and attenuation correction based on the reference CT or pseudo CT
μ-maps. The PET image reconstructed using the reference CT-based
μ-map was considered as the reference PET. An Ordered Subsets Expectation Maximisation (OSEM) algorithm with 3 iterations of 21 subsets was used. Effects of PSF and randoms were not included and post-reconstruction smoothing was not applied. The simulation and reconstruction were performed using STIR [
29].
Algorithmic comparison
The performance of the proposed synthesis algorithm was compared with ground truth data for 22 subjects following a leave-one-out cross-validation scheme. For each subject, and both PET tracers, 5 pseudo CTs were synthesised using a database of 21 subjects following the method from [
17] and the improvements proposed. An additional pseudo CT was synthesised using the SPM-based approach [
18]:
-
pCT
SPM using the method described in [
18] given the T1 image;
-
pCT
T1 using the method described in [
17] given the T1 image;
-
pCT
T2 using the method described in [
17] given the T2 image;
-
npCTT1 using the ROI-LNCC on the T1 image;
-
npCTT2 using the ROI-LNCC on the T2 image;
-
npCTT1,T2 using the MV-ROI-LNCC on the T1-T2 pair of images.
In order to preserve the alignment of the CT and PET images, the two PET/CT acquisitions were considered independently, meaning that we created two MR-CT databases: one with the CT images from the 18F-FDG PET/CT scan and the MR images; and another one with the CT images from the 18F-florbetapir PET/CT scan and the MR images.
The quantitative validation consisted of two steps:
1.
The pseudo CTs were compared to the subject’s original CT image, validating the accuracy of the CT synthesis.
2.
For both FDG and florbetapir tracers, the simulated PET data were reconstructed using the different pCT μ-maps, and compared with the reference PET reconstructed using the CT-based μ-map, validating the accuracy of the PET attenuation correction.
Statistical significance was assessed using the paired one-tailed Wilcoxon signed-rank test, with a 5% significance level.
Pseudo CT accuracy
For every subject, the mean absolute error and the mean error, defined respectively as
$$\begin{array}{@{}rcl@{}} \text{MAE}&=&\frac{1}{N_{V}} {\sum\limits}_{v\in V}|I_{v}-R_{v}| \;\;\;\text{and} \\ \text{ME}&=&\frac{1}{N_{V}} {\sum\limits}_{v\in V}{\left( I_{v}-R_{v}\right)} , \end{array} $$
(1)
were calculated between the reference CT (
R =
C
T) and each of the pseudo CTs (
I =
p
C
T), in a region of interest
V comprising
N
V
voxels. This region of interest is limited to a mask defined by segmenting the head from the background using the original CT. The MAE provides information about reconstruction error and deviations from the expected values while the ME gives information about an inherent bias in the methodology.
To localise the error and bias introduced by each approach, the original CTs and pseudo CTs from the 22 subjects were mapped to a common space via a groupwise registration [
30]. Difference maps were then computed between the original CT and the pseudo CTs and averaged across all the subjects. Results are presented using a mean intensity projection: the mean value along a projection line is assigned to the pixel represented on the projected image.
PET accuracy
We first computed the relative MAE and ME in the reference regions, defined respectively as
$$\begin{array}{@{}rcl@{}} \text{rMAE}&=&100*\frac{{\sum\limits}_{v\in V}{|I_{v}-R_{v}|}}{{\sum}_{v\in V}{R_{v}}}\;\;\;\; \text{and} \\ \text{rME}&=&100*\frac{{\sum\limits}_{v\in V}{\left( I_{v}-R_{v}\right)}}{{\sum}_{v\in V}{R_{v}}} \;\;\;, \end{array} $$
(2)
between the reference PET (
R =
P
E
T
C
T,r
e
f
) and each of the PETs corrected with the synthetic
μ-maps (
I =
P
E
T
p
C
T,r
e
f
). This analysis aims at characterising the presence of error and bias in the reference regions.
In order to provide a quantitative regional assessment of the error and bias after normalisation by the reference region, we computed the relative MAE and ME between the normalised reference PET (R = P
E
T
C
T
/m
e
a
n
r
e
f
) and each of the PETs corrected with the synthetic μ-maps (I = P
E
T
p
C
T
/m
e
a
n
r
e
f
). To assess the performance of the proposed method in areas relevant to dementia pathologies, we analysed results in the full brain region, but also in the posterior cingulate gyrus, angular gyrus, superior frontal gyrus, fusiform gyrus and anterior cingulate gyrus.
Finally, the PET images from the 22 subjects were mapped to a common space via a CT-based groupwise registration [
30]. Difference maps were then computed between the reference PET and each of the PETs corrected with the pCT
μ-maps, and their average and standard deviation, across all the subjects, displayed using a mean intensity projection.
Discussion
Correcting for attenuation is an essential requirement to perform an accurate quantitative analysis of PET data. In this paper, we jointly validated an MR-based attenuation correction method on two different tracers using 22 subjects with dementia. Each subject underwent two PET/CT scans, with 18F-FDG and 18F-florbetapir tracers, providing reference CT and PET images, and an MRI scan providing T1-weighted and T2-weighted images. The evaluation was conducted both in the full brain and in ROIs relevant for dementia (posterior cingulate gyrus, angular gyrus, superior frontal gyrus, fusiform gyrus and anterior cingulate gyrus).
While validating the method using the pseudo CTs from [
17], we noticed problems in locations close to the edges of the template database field-of-view, such as the areas surrounding the cerebellum. This was problematic because the cerebellum was used as a SUVR normalising region. By introducing some methodological improvements, we were able to improve the CT synthesis accuracy. First, we proposed a new similarity measure for irregular regions-of-interest (ROI-LNCC) to increase the accuracy of the synthesis at the borders of the field-of-view. As a second step, we extended the method from [
17] to multi-contrast MR data, allowing the introduction of complementary information. We showed that one can synthesise CT from T1 images, T2 images, but also combinations of MR contrasts. Combining complementary information describing the underlying subject’s anatomy, at both the registration and image similarity stages, reduces the ill-posedness of the problem. As a consequence, the CT synthesis error decreases, which leads to reduced errors when comparing PET images to the reference PETs. By combining these two improvements, the average MAE in the full brain for the florbetapir PET images decreased from 2.56% to 1.60%, with a reduction in MAE variance from 1.17% to 0.41% (Table
2).
When analysing the results of the CT synthesis, we note that the errors are mostly located in the sinus area and at the bone/dura/CSF boundaries (Fig.
2), regions with a low tissue discriminative power. We demonstrated that, for both tracers, in the brain region, the error when comparing PET images corrected with the pseudo CTs to the reference PET images corrected using the original CT is less than 2% (Table
2). While in ROIs close to the skull (superior frontal gyrus, angular gyrus and fusiform gyrus), the MAE can be up to 3%, in deeper structures (posterior cingulate gyrus, and anterior cingulate gyrus), the MAE is below 1.2%. The variance in SUVR explained by the attenuation correction error is likely to be smaller than the intrinsic PET noise variance [
31].
We also compared our approach to a state-of-the-art, template-based, method using SPM. The pseudo CT is generated by non-rigidly registering the segmented patient’s T1 image to a segmented T1 template and by applying the inverse transformation to the associated CT template. Generating the MRI and CT templates through a co-registration and averaging process leads to smooth pseudo CT images. Moreover, the outcome of the process depends on a single registration, which might be inaccurate. As a result, for both PET tracers, the proposed method outperforms the SPM-based approach.
To the best of our knowledge, this is the first time that an MR-based attenuation correction method has been jointly validated with two different PET tracers. While the FDG PET images were only moderately affected by the inaccuracies at the border of the FOV of the original pseudo CTs [
17], these inaccuracies had a major impact on the florbetapir PET images as they were normalised using the cerebellar grey matter. With the improvements proposed, CT images were more accurately synthesised at the borders of the FOV, which includes the cerebellum area, thus reducing the errors observed in the florbetapir PET images. Our validation showed that, for both tracers, the PET images were accurately corrected for attenuation.
The proposed method does not require the acquisition of PET analysis specific MRI sequences, such as Dixon or UTE, which means that the acquisition protocol can be entirely dedicated to clinically-relevant sequences. In this paper, we used T1 and T2 images as these sequences are usually acquired as part of many standard acquisition protocols, but the method could be extended to any combination of sequences providing enough structural information and structural contrast. The resolution of both the T1 and T2 images was high (1.1 mm isotropic), which might not always be the case in clinical practice. For example, if the resolution of the T2 image is lower than the resolution of the T1 image, its contribution to the synthesis process may be limited. The quality of the results obtained is likely to be between the quality reached when using two high-resolution images and the quality reached when only using one sequence.
The scope of applications of the proposed methodology exceeds the field of PET/MR. For example, μ-map synthesis methods can also be used to correct the PET images for attenuation when the radiation dose needs to be kept to a minimum, such as for paediatric subjects. Using an appropriate database, a pseudo CT could be synthesised from an MRI image acquired during a previous examination, then registered to the non corrected PET image and finally used to correct the PET data.
While we focused our work on brain applications using subjects who do not present unusual or highly abnormal skull anatomies, further experiments are required to validate the method on subjects with pathologies affecting regions critical for attenuation correction, such as the skull, and in other regions of the body. As long as the morphological variability is represented in the database and the registration between MRI pairs is sufficiently accurate, the technique could, in theory, be applied to other body parts. Furthermore, the inclusion of clinical information (patient’s gender, age, weight or ethnicity), as suggested by [
32], could be used to improve the bone-density estimates.