G Model
ARTICLE IN PRESS
EURR-6724; No. of Pages 7
European Journal of Radiology xxx (2014) xxx–xxx
Contents lists available at ScienceDirect
European Journal of Radiology journal homepage: www.elsevier.com/locate/ejrad
Whole-body PET/MRI: The effect of bone attenuation during MR-based attenuation correction in oncology imaging M.C. Aznar a,∗ , R. Sersar b , J. Saabye b , C.N. Ladefoged c , F.L. Andersen c , J.H. Rasmussen a , J. Löfgren c , T. Beyer d a
Department of Oncology, Section of Radiotherapy 3994, Rigshospitalet, Blegdamsvej 9, 2100 Copenhagen, Denmark DTU Informatics, Technical University of Denmark, Kongens Lyngby, Denmark c Department of Clinical Physiology, Nuclear Medicine and PET, Rigshospitalet, Copenhagen, Denmark d Centre for Medical Physics and Biomedical Engineering, Medical University of Vienna, Vienna, Austria b
a r t i c l e
i n f o
Article history: Received 6 February 2014 Received in revised form 13 March 2014 Accepted 17 March 2014 Keywords: Combined PET/MR Attenuation correction Bone Quantification
a b s t r a c t Purpose: In combined PET/MRI standard PET attenuation correction (AC) is based on tissue segmentation following dedicated MR sequencing and, typically, bone tissue is not represented. We evaluate PET quantification in whole-body (WB)-PET/MRI following MR-AC without considering bone attenuation and then investigate different strategies to account for bone tissue in clinical PET/MR imaging. To this purpose, bone tissue representation was extracted from separate CT images, and different bone representations were simulated from hypothetically derived MR-based bone classifications. Methods: Twenty oncology patients referred for a PET/CT were injected with either [18F]-FDG or [18F]-NaF and imaged on PET/CT (Biograph TruePoint/mCT, Siemens) and PET/MRI (mMR, Siemens) following a standard single-injection, dual-imaging clinical WB-protocol. Routine MR-AC was based on in/opposed-phase MR imaging (orgMR-AC). PET(/MRI) images were reconstructed (AW-OSEM, 3 iterations, 21 subsets, 4 mm Gaussian) following routine MR-AC and MR-AC based on four modified attenuation maps. These modified attenuation maps were created for each patient by non-linear co-registration of the CT images to the orgMR-AC images, and adding CT bone mask values representing cortical bone: 1200 HU (cortCT), spongiosa bone: 350 HU (spongCT), average CT value (meanCT) and original CT values (orgCT). Relative difference images of the PET following AC using the modified attenuation maps were compared. SUVmean was calculated in anatomical reference regions and for PET-positive lesions. Results: The relative differences in SUVmean across patients following orgMR-AC and orgCT in soft tissue lesions and in bone lesions were similar (range: 0.0% to −22.5%), with an average underestimation of SUVmean of 7.2% and 10.0%, respectively when using orgMR-AC. In bone lesions, spongCT values were closest to orgCT (median bias of 1.3%, range: –9.0% to 13.5%) while the overestimation of SUVmean with respect to orgCT was highest for cortCT (40.8%, range: 1.5% to 110.8%). For soft tissue lesions the bias was highest using cortCT (13.4%, range: –2.3% to 17.3%) and lowest for spongCT (–2.2%, range: 0.0% to –13.7%). Conclusions: In PET/MR imaging using standard MR-AC PET uptake values in soft lesions and bone lesions are underestimated by about 10%. In individual patients this bias can be as high as 22%, which is significant during clinical follow-up exams. If bone segmentation is available, then assigning a fixed attenuation value of spongious bone to all bone structures appears reasonable and results in only a minor bias of 5%, or less in uptake values of soft tissue and bone lesions. © 2014 Published by Elsevier Ireland Ltd.
1. Introduction
∗ Corresponding author. Tel.: +45 35 45 4830; fax: +45 35 45 3990. E-mail addresses: [email protected] (M.C. Aznar), [email protected] (R. Sersar), julie [email protected] (J. Saabye), [email protected] (C.N. Ladefoged), [email protected] (F.L. Andersen), [email protected] (J.H. Rasmussen), [email protected] (J. Löfgren), [email protected] (T. Beyer).
Dual-modality PET/CT imaging has become a modality-ofchoice for the diagnostic work-up in oncology patients [1]. Combined PET/CT imaging provides fully-aligned PET and CT images over extended co-planar imaging ranges as well as quantitative PET information following CT-based attenuation and scatter correction [2]. Recently, combined PET/MRI imaging systems have
http://dx.doi.org/10.1016/j.ejrad.2014.03.022 0720-048X/© 2014 Published by Elsevier Ireland Ltd.
Please cite this article in press as: Aznar MC, et al. Whole-body PET/MRI: The effect of bone attenuation during MR-based attenuation correction in oncology imaging. Eur J Radiol (2014), http://dx.doi.org/10.1016/j.ejrad.2014.03.022
G Model EURR-6724; No. of Pages 7 2
ARTICLE IN PRESS M.C. Aznar et al. / European Journal of Radiology xxx (2014) xxx–xxx
been proposed as an alternative to PET/CT [3,4], whereby the increased soft tissue contrast through the MR and the overall reduction in patient exposure, primarily from the lack of ionizing radiation exposure from the CT, are considered key drivers for this technology [5]. Today, three different design concepts exist for PET/MRI [6–8] of which the fully-integrated concept [6] offers the ability to acquire PET and MRI data simultaneously, in addition to the advantages described above. Following the introduction of the first commercial PET/MRI system in 2010, clinical validation of these systems is still ongoing [5], and several technical challenges mandate further investigation [5,9]. One of the key challenges of combined PET/MRI, next to the modifications of the system hardware, is the lack of a CT-like transmission source and the subsequent need to perform attenuation and scatter correction using the available MR and emission information only. In the co-planar design, for example, MR-based attenuation correction (MR-AC) is performed using a dedicated T1weighted MR sequence in combination with a customized image processing technique to derive attenuation maps for whole-body PET [7]. The resulting MR-based attenuation map is composed of three classes of voxels: air (outside the subject), lungs and soft tissue. In the integrated design MR-AC is performed using an in-/opposed-phase MR sequence followed by an image-based segmentation [10] into four tissue classes: air (outside the subject), lungs, soft tissue and fat. Of note, this approach does not account for the presence of bone in the attenuation map. The absence of bone, however, could lead to a noticeable overall and local underestimation of the attenuation correction factors and, therefore, to a bias of the AC-PET data. Martinez-Möller et al. [11] simulated the absence of bone attenuation using PET/CT data from 35 oncology patients and concluded that the bias of AC-PET, measured as relative changes in standardized uptake values (SUV), was highest for lesions in the pelvic bone with an average SUV underestimation of 13%. The mean SUV for bone lesions across all patients decreased by 8% when using the segmented attenuation map without accounting for bone. Samarin and colleagues studied the effect of ignoring bone attenuation during MR-AC further [12]. They report an average decrease in SUV in bony structures and soft tissue lesions adjacent to bone of 11% (range: 2% to 31%) and 3% (0% to 4%), respectively relative to CT-based AC. The authors also demonstrate a dependency of the SUV bias on the composition of the bone lesions. Despite a general underestimation of the SUV in combined whole-body PET/MRI images none of the studies above reported on missing out on bone lesions entirely from standard MR-based AC [13]. Nonetheless, full quantitative accuracy is required for reliable therapy response assessment and further work to ensure proper representation of the bones is warranted. New developments suggest that bone segmentation based solely on MR images should be possible, at least for the brain region when employing UTE sequences [14] or atlas-based approaches [15]. However, in either case, assigning a relevant and accurate density distribution to the segmented bone structures is challenging. The goal of this study was to assess the effect of MR-AC on the accuracy and visibility of PET positive lesions in PET images of oncology patients undergoing integrated PET/MRI imaging. Here, we assume the availability of MR image volumes that represent particular bone structures as acquired with hypothetical MR sequences. Corresponding MR-based attenuation maps are derived from standard MR-AC maps following the 4-class tissue mode (background, lung tissue, fat, soft tissue) [11] with several segmented bone attenuation maps superimposed from separate CT scans, thereby assuming that such bone maps will become available through novel MR acquisition and processing in the future. We assessed SUV in reference tissues and lesions of PET/MRI images of oncology patients with lesions close to or within bony structures
following MR-AC using the original MR-based attenuation map and the modified attenuation maps. 2. Materials and methods 2.1. Patients Twenty clinical oncology patients with a referral for a PET/CT study were included in this study. Of those 10 patients with soft tissue lesions in the head/neck region were scheduled for an [18F]FDG examination (Group A) and 10 patients with osseous lesions were scheduled for a whole-body [18F]-NaF PET/CT examination (Group B). All patients underwent combined PET/CT and PET/MRI imaging between February 2012 and February 2013. 2.2. Imaging protocol The study protocol corresponded to a single-injection, dualimaging protocol and was approved by the local ethics committee. All patients underwent PET/CT imaging (Biograph TruePoint/mCT, Siemens) first. PET/CT imaging was performed without time-offlight (TOF) and images were reconstructed without using spatial resolution recovery [16], as this option was not available on subsequent PET/MRI. Patients in Group A were injected with 306 MBq ± 80 MBq of [18F]-FDG. Patients were positioned with their arms down and fixed with a thermoplastic mask over the head/neck region. PET/CT imaging started at 62 min ± 4 min post-injection. First, a topogram scan was acquired for the delineation of the co-axial imaging range, followed by a spiral CT scan (120 kV p, approximately 200 effective mAs adjusted after Siemens CareDose) and a 2-bed emission acquisition (2–3 min/bed). Patients in Group B were injected with 203 MBq ± 21 MBq of [18F]-NaF. Patients were positioned with their arms down. PET/CT imaging started at 49 min ± 5 min post-injection. First, a topogram scan was acquired for the delineation of the co-axial imaging range, followed by a spiral CT scan (120 kV p, 40 mA s) and a multi-bed (four to seven bed positions) emission acquisition (2 min/bed). Following the PET/CT examination patients were repositioned on the integrated PET/MRI system (Biograph mMR, Siemens). All patients were positioned head-first supine and with their arms down. In head/neck patients the thermoplastic mask was not used but the neck was supported by a radiotherapy cushion to provide similar positioning of the head/neck region, while allowing the use of the head/neck coil for MR. PET/MRI imaging started at 117 min ± 27 min post-injection for 20 min per bed position for patients in Group A and 130 min ± 32 min post-injection for 5 min per bed position for patients in Group B. The clinical imaging protocol included a scout scan for defining the co-axial imaging range, in-phase and opposed-phase MR imaging, using the Dixon VIBE sequence, for deriving MR-based attenuation factors (orgMR-AC) [11] as well as simultaneously acquired emission data and T1/T2 MR acquisitions. Protocol details are given in [17]. For the purpose of this study, PET data from the PET/MRI acquisition were corrected for attenuation using the orgMR-AC method as well as attenuation maps derived from co-registered CT data from the prior PET/CT examination. 2.3. Co-registration and bone segmentation All co-registration and segmentation procedures were performed in MATLAB (version 2011b, MathWorks, USA) and DICOM files were converted using the CERR package [18]. The CT images from the PET/CT examination were co-registered to the MR attenuation map (orgMR-AC) using a rigid transformation followed by a non-rigid transformation (cubic b-spline). All
Please cite this article in press as: Aznar MC, et al. Whole-body PET/MRI: The effect of bone attenuation during MR-based attenuation correction in oncology imaging. Eur J Radiol (2014), http://dx.doi.org/10.1016/j.ejrad.2014.03.022
G Model EURR-6724; No. of Pages 7
ARTICLE IN PRESS M.C. Aznar et al. / European Journal of Radiology xxx (2014) xxx–xxx
3
Fig. 1. Derivation of the attenuation maps with different representations of bone. Bone values were derived from the available CT image data: original HU values for bone (“orgCT”), average of the original HU values (“meanCT”) of the segmented CT volume, fixed value of 350 HU (“spongCT”) and fixed value of 1200 HU (“cortCT”). orgMC-AC refers to the original in-/opposed-phase attenuation map (no bones present).
co-registration steps were performed using MATLAB using a 4level normalized cross-correlation and a regularization parameter of 10−8 [19]. After the co-registration step, bone structures were extracted from the available CT images using a k-means clustering algorithm, which clustered the image intensities into four classes (air, soft tissue, fat and bone) and defined a centroid value representation the mean intensity of the bones. This centroid value was then used as input for a Potts model to assign each voxel to a specific class by including information about neighbouring voxels. As a last step, a morphological closing operation was performed to remove potential gaps and holes within the segmented skeleton.
2.4. CT-derived attenuation maps Four different strategies were investigated regarding the representation of bone attenuation values and subsequent attenuation correction (Fig. 1). The first attenuation map (orgCT) was based on the co-registered original CT attenuation values (HU). For the second attenuation map (meanCT) all values in the segmented bone contour were assigned the mean CT attenuation (HU) of the original bone voxels, simulating the attribution of bone density after a template-based segmentation [15]. Mean CT attenuation values were calculated per patient. The third and fourth attenuation maps employed the coregistered CT with the bone tissue values being replaced by fixed
Fig. 2. Definition of regions and volumes of interest for SUVmean: (A) for head and neck patients (Group A) and (B) for bone lesion patients (Group B). Reference regions and PET-positive lesions are shown in blue and red, respectively. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)
Please cite this article in press as: Aznar MC, et al. Whole-body PET/MRI: The effect of bone attenuation during MR-based attenuation correction in oncology imaging. Eur J Radiol (2014), http://dx.doi.org/10.1016/j.ejrad.2014.03.022
G Model EURR-6724; No. of Pages 7 4
ARTICLE IN PRESS M.C. Aznar et al. / European Journal of Radiology xxx (2014) xxx–xxx
Fig. 3. Case study for head and neck patients (Group A). Top row: series of attenuation maps according to Fig. 1 (converted to HU); bottom row: AC-PET images (in counts). Note, markedly different uptake in the bone region but little relative differences in the soft tissue lesion (arrow).
CT attenuation values for spongiosa and cortical bone, respectively, and simulating the attribution of bone density following a UTEbased segmentation [14]. The attenuation values for spongious and cortical bone were set to 350 HU and 1200 HU independent of the patient [20]. The resulting attenuation maps were named spongCT and cortCT (Fig. 1). Each of the three alternative approaches to MR-AC (meanCT, spongCT and cortCT) corresponds to a clinical situation where bone structures would be classified on dedicated MR acquisitions, e.g., [14], while ignoring intra-tissue variations of physical bone density. Attenuation values at 511 keV were calculated for each voxel in the four attenuation maps above by applying the segmentation, bi-linear scaling approach by Carney et al. [21]. The modified maps (i.e., including the segmented bones from CT) were inserted in the original MR-AC DICOM files replacing the original values. The modified DICOM files were then imported into the reconstruction computer according to instructions provided by the vendor. PET/MRI images were reconstructed following AC using orgMRAC and any of the four derived attenuation maps above. Image reconstruction was based on 3D-OSEM with 3 iterations and 21 subsets on 344 by 344 matrices with a 4 mm Gaussian filter applied.
2.5. Image analysis We assessed the relative changes of SUV in reference regions and lesions of PET/MRI images following AC using any of the three modified attenuation maps meanCT, spongCT and cortCT as well as the PET images following AC using standard MR-AC (orgMRAC) and the reference orgCT. Fig. 2 summarizes the location of the regions (ROI and VOI) used for the region-based analysis of PET images of patients in Group A and B. All regions were defined manually by an experienced nuclear medicine physician on the PET/MRI images following orgMR-AC using the Mirada software (version XD3.4, Mirada Medical Ltd., UK). In Group A (Fig. 2A), reference regions were defined in the upper spine and in the cerebellum using an ROI in the axial plane and a spherical VOI, respectively. Uptake within the lesion was calculated from an isocontour at 40% of SUVmax. In the bone lesions patients (Group B), reference regions (see Fig. 2B) were defined as the bladder (spherical VOI in the centre of the organ), the femoral head (spherical VOI), the spine (ROI in the axial plane) and the pelvic bone (ROI in the coronal plane). Similar to Group A, uptake within the lesion was calculated from an isocontour at 40% of SUVmax.
Fig. 4. Relative difference (%) of average SUVmean in reference regions and lesions for head and neck patients in Group A. SB—spinal bone.
Please cite this article in press as: Aznar MC, et al. Whole-body PET/MRI: The effect of bone attenuation during MR-based attenuation correction in oncology imaging. Eur J Radiol (2014), http://dx.doi.org/10.1016/j.ejrad.2014.03.022
G Model EURR-6724; No. of Pages 7
ARTICLE IN PRESS
19.8 (6.8; 63.2) 1.3 (−9.0; 13.5) 15.0 (5.8; 55.4) 17.7 (6.8; 57.6) 12.5 (5.4; 52.7) 13.9 (5.7; 57.7)
−10.0 (−22.4; 0.0)
16.5 (−4.5; 55.3)
RD (%) cortCT
12.6 (2.5; 26.0) −2.2 (−13.7; 0.0)
RD (%) spongCT
10.8 (2.2; 23.0) −7.2 (−22.5; 0.0)
RD (%) meanCT
12.8 (2.4; 25.3)
RD (%)
9.9 (2.2; 21.8)
orgMR-AC
Patients in Group A included 3 women and 7 men, with an average age of 62 years (range 53;72) and an average weight of 76 kg (range 38;123). Fig. 3 shows a case study of a head/neck patient from Group A processed with the four different attenuation maps (Fig. 1). The PET-positive lesion was visible across the different PET images. Table 1 summarizes the relative differences in SUVmean for all soft tissue lesions in patients from Group A with respect to values from orgCT. In all but one patient, orgMR-AC caused an underestimation of the SUVmean value with respect to orgCT (see Table S1, Supplementary material). In contrast, the other CTbased AC method tended to lead to an overestimation of SUVmean values. There was a large variation of apparent lesion uptake across patients. The differences in SUVmean between orgMR-AC and orgCT ranged from 0.0% to −22.5% (median −7.2%). The maximum median difference (12.6%) was observed between orgCT and cortCT though meanCT could result in as high a bias as cortCT in individual patients (see Table 1). Fig. 4 summarizes the relative changes in SUVmean averaged across all head/neck patients in Group A for the reference regions and soft tissue lesions. Relative differences were smaller in the cerebellum, on average −14% (range −17% to −8%) for SUVmean between orgMR-AC and orgCT. The upper spine is the region most affected by variations in AC with a relative difference in SUVmean between orgMR-AC and orgCT of −27% across patients (range: −40% to −8%). This relative difference was as high as 165% for cortCT in the upper spine. Patients in Group B included 4 women and 6 men, with an average age of 62 years (range 34;75) and an average weight of 84 kg (range 68;110). Fig. 5 shows a case study of a patient with bone lesions from Group B processed with the four attenuation maps (Fig. 1). The PET-positive lesion was seen on all PET images but exaggerated on cortCT and spongCT. Table 1 summarizes also the relative differences in SUVmean for the bone lesions in each patient of Group B with respect to values from orgCT. Relative differences of SUVmean of the lesions were strongly affected by the different AC maps: the maximum relative difference ranges from −22.4% with orgMR-AC to 110.8% with cortCT. On average, values obtained with spongCT displayed the smallest bias with respect to orgCT (1.3%). Fig. 6 summarizes the relative differences in SUVmean of reference regions and lesions in patients in Group B. The largest difference with respect to orgCT was observed for cortCT for all regions. The bladder was visibly unaffected by the differences in the AC maps, with an average bias of 2% for SUVmean between orgMR-AC and orgCT (the maximum bias observed in an
orgCT
3. Results
11.3 (2.2; 23.5)
with SUVrec being the SUVmean for a given attenuation map (orgMR-AC, meanCT, spongCT or cortCT)) and SUVorgCT being the corresponding SUVmean in PET images following orgCT, considered to be the gold standard. All SUV were related to the patient body weight. Relative differences were reported for each reference region and lesion. Here, reference region referred to an area of physiologic tracer uptake in an organ (e.g., bladder) or in a substructure of an organ (e.g., bone) as illustrated in Fig. 2. We calculated the mean relative difference (%) as the average across all 10 patients in Group A and all 10 patients in Group B for reference regions, bone lesions and soft tissue lesions.
Head and neck lesions median (range) Bone lesions median (range)
(1)
SUVmean
SUVrec − SUVorgCT SUVorgCT
Table 1 SUVmean values of head and neck lesions (10 patients, Group A) and bone lesions (10 patients, Group B). Results are displayed as median and range. All relative differences are reported with respect to orgCT.
RD% = 100 ×
10.8 (1.6; 17.3)
For each of the four attenuation maps, relative differences (%) of attenuation-corrected PET images were assessed for all reference regions and lesions:
40.8 (1.5; 110.8)
5
13.4 (−2.3; 17.3)
M.C. Aznar et al. / European Journal of Radiology xxx (2014) xxx–xxx
Please cite this article in press as: Aznar MC, et al. Whole-body PET/MRI: The effect of bone attenuation during MR-based attenuation correction in oncology imaging. Eur J Radiol (2014), http://dx.doi.org/10.1016/j.ejrad.2014.03.022
G Model EURR-6724; No. of Pages 7
ARTICLE IN PRESS M.C. Aznar et al. / European Journal of Radiology xxx (2014) xxx–xxx
6
Fig. 5. Case study for bone lesion patients (Group B). Top row: series of attenuation maps according to Fig. 1 (converted to HU); bottom row: AC-PET images (in counts). Case includes a bone lesion in the left pelvis (arrow).
individual patient: 20% with cortCT) while the femoral head bone was the reference region most affected.
4. Discussion In this study, we assessed the effect of using different attenuation maps on PET image quality and accuracy of PET/MRI data of oncology patients. Our first objective was similar to that of previous studies by Samarin [12] and Martinez-Möller [11]. However, here we use data from fully-integrated PET/MRI. Issues such as potential PET/MR-specific artifacts and correction for the presence of MR coils in the PET attenuation do not mandate special attention as they are independent of the relative measures (Eq (1)) derived from alterations of the patient-based attenuation maps.
In our patient cohort, we observed a similar bias in soft tissue lesions close to the bone and in bone lesions of −7.2% and −10% (range 0%; −22%). Here, we assumed orgCT to be the gold standard. While this value supports previous reports for the bias in bone lesions [11,12], it is slightly higher than those reported by Samarin [12] for soft tissue lesions. In their study they reported a bias of −3.2% (range 0%; −4%), though the histology in their patient group was not specified. In our cohort, all soft tissue lesions were squamous cell carcinoma of the head and neck. Large tumours and smaller positive lymph nodes were present in this cohort and one cannot exclude that a potential bias as high as 22% might cause a small positive lesion to be undetected. Moreover, uptake in reference bone regions was also affected by not accounting for bone during MR-AC. In our patient cohort, this bias was most pronounced in the upper spine of patients in Group
Fig. 6. Relative difference (%) of average SUVmean in reference regions and lesions for head and neck patients in Group B. Femoral HB—femoral head bone.
Please cite this article in press as: Aznar MC, et al. Whole-body PET/MRI: The effect of bone attenuation during MR-based attenuation correction in oncology imaging. Eur J Radiol (2014), http://dx.doi.org/10.1016/j.ejrad.2014.03.022
G Model EURR-6724; No. of Pages 7
ARTICLE IN PRESS M.C. Aznar et al. / European Journal of Radiology xxx (2014) xxx–xxx
A. In contrast the cerebellum and especially the bladder were relatively unaffected in patients in Group A and Group B, respectively. Our second objective was to investigate which density to assign to the segmented bone structures using MR-attenuation maps corresponding to variations of existing as well as hypothetical approaches to MR-AC. For example, orgCT, would correspond to an attenuation map derived from MR with the full anatomical variability of attenuation values of soft tissue and bone for which no standard MR-only method is available today. The attenuation data in spongCT and cortCT would represent soft tissue values as derived from in/opposed phase imaging merged with a segmented bone representation (e.g., from an UTE sequence) whereby all bone values are set to fixed attenuation values representing spongiosa bone or cortical bone. Using the information from a separate CT for segmentation is a time-consuming procedure and is not an option for routine clinical adoption but reconstructing PET images with simple in/opposed phase imaging (orgMR-AC) already results in a bias in both bone lesion and soft tissue lesion uptake. Our results suggest that assigning a uniform density corresponding to spongious bone is a reasonable compromise. Indeed, in bone lesions, the overestimation of SUVmean with respect to orgCT was highest for cortCT (40.8%, range: 1.5% to 110.8%) while spongCT values were the closest to orgCT (average bias of 1.3%, range: −9.0 to 13.5%). For soft tissue lesions the bias was highest with cortCT (13.4%, range: −2–17%) but lowest with spongCT (−2−2%, range 0.0% to −13.7%). In average across all patients, the bias in SUVmean between spongCT and orgCT was less than 5% for lesions in Group A and in Group B. This strategy could provide a realistic alternative to a separate CT acquisition assuming that the bones can be accurately segmented. A limitation of our study is that our procedure is sensitive to co-registration and segmentation errors, mainly due to positioning differences between the CT and the PET/MRI examination. However, on visual examination, patient position was very reproducible between PET/CT and PET/MR and, thus, residual misalignment was not considered to be a major source of error. Our results suggest that an accurate representation of the bones is needed to assure the quantitative accuracy of PET/MRI data. In the absence of bone attenuation factors, the accuracy of the quantification of bone lesions, as well as soft tissue lesions relatively close to the bone, is inaccurate. If bone segmentation as obtained via a template-based approach or from dedicated UTE sequences is available, then assigning a fixed attenuation value of spongious bone to the whole skeleton results in only a minor bias (
ARTICLE IN PRESS
EURR-6724; No. of Pages 7
European Journal of Radiology xxx (2014) xxx–xxx
Contents lists available at ScienceDirect
European Journal of Radiology journal homepage: www.elsevier.com/locate/ejrad
Whole-body PET/MRI: The effect of bone attenuation during MR-based attenuation correction in oncology imaging M.C. Aznar a,∗ , R. Sersar b , J. Saabye b , C.N. Ladefoged c , F.L. Andersen c , J.H. Rasmussen a , J. Löfgren c , T. Beyer d a
Department of Oncology, Section of Radiotherapy 3994, Rigshospitalet, Blegdamsvej 9, 2100 Copenhagen, Denmark DTU Informatics, Technical University of Denmark, Kongens Lyngby, Denmark c Department of Clinical Physiology, Nuclear Medicine and PET, Rigshospitalet, Copenhagen, Denmark d Centre for Medical Physics and Biomedical Engineering, Medical University of Vienna, Vienna, Austria b
a r t i c l e
i n f o
Article history: Received 6 February 2014 Received in revised form 13 March 2014 Accepted 17 March 2014 Keywords: Combined PET/MR Attenuation correction Bone Quantification
a b s t r a c t Purpose: In combined PET/MRI standard PET attenuation correction (AC) is based on tissue segmentation following dedicated MR sequencing and, typically, bone tissue is not represented. We evaluate PET quantification in whole-body (WB)-PET/MRI following MR-AC without considering bone attenuation and then investigate different strategies to account for bone tissue in clinical PET/MR imaging. To this purpose, bone tissue representation was extracted from separate CT images, and different bone representations were simulated from hypothetically derived MR-based bone classifications. Methods: Twenty oncology patients referred for a PET/CT were injected with either [18F]-FDG or [18F]-NaF and imaged on PET/CT (Biograph TruePoint/mCT, Siemens) and PET/MRI (mMR, Siemens) following a standard single-injection, dual-imaging clinical WB-protocol. Routine MR-AC was based on in/opposed-phase MR imaging (orgMR-AC). PET(/MRI) images were reconstructed (AW-OSEM, 3 iterations, 21 subsets, 4 mm Gaussian) following routine MR-AC and MR-AC based on four modified attenuation maps. These modified attenuation maps were created for each patient by non-linear co-registration of the CT images to the orgMR-AC images, and adding CT bone mask values representing cortical bone: 1200 HU (cortCT), spongiosa bone: 350 HU (spongCT), average CT value (meanCT) and original CT values (orgCT). Relative difference images of the PET following AC using the modified attenuation maps were compared. SUVmean was calculated in anatomical reference regions and for PET-positive lesions. Results: The relative differences in SUVmean across patients following orgMR-AC and orgCT in soft tissue lesions and in bone lesions were similar (range: 0.0% to −22.5%), with an average underestimation of SUVmean of 7.2% and 10.0%, respectively when using orgMR-AC. In bone lesions, spongCT values were closest to orgCT (median bias of 1.3%, range: –9.0% to 13.5%) while the overestimation of SUVmean with respect to orgCT was highest for cortCT (40.8%, range: 1.5% to 110.8%). For soft tissue lesions the bias was highest using cortCT (13.4%, range: –2.3% to 17.3%) and lowest for spongCT (–2.2%, range: 0.0% to –13.7%). Conclusions: In PET/MR imaging using standard MR-AC PET uptake values in soft lesions and bone lesions are underestimated by about 10%. In individual patients this bias can be as high as 22%, which is significant during clinical follow-up exams. If bone segmentation is available, then assigning a fixed attenuation value of spongious bone to all bone structures appears reasonable and results in only a minor bias of 5%, or less in uptake values of soft tissue and bone lesions. © 2014 Published by Elsevier Ireland Ltd.
1. Introduction
∗ Corresponding author. Tel.: +45 35 45 4830; fax: +45 35 45 3990. E-mail addresses: [email protected] (M.C. Aznar), [email protected] (R. Sersar), julie [email protected] (J. Saabye), [email protected] (C.N. Ladefoged), [email protected] (F.L. Andersen), [email protected] (J.H. Rasmussen), [email protected] (J. Löfgren), [email protected] (T. Beyer).
Dual-modality PET/CT imaging has become a modality-ofchoice for the diagnostic work-up in oncology patients [1]. Combined PET/CT imaging provides fully-aligned PET and CT images over extended co-planar imaging ranges as well as quantitative PET information following CT-based attenuation and scatter correction [2]. Recently, combined PET/MRI imaging systems have
http://dx.doi.org/10.1016/j.ejrad.2014.03.022 0720-048X/© 2014 Published by Elsevier Ireland Ltd.
Please cite this article in press as: Aznar MC, et al. Whole-body PET/MRI: The effect of bone attenuation during MR-based attenuation correction in oncology imaging. Eur J Radiol (2014), http://dx.doi.org/10.1016/j.ejrad.2014.03.022
G Model EURR-6724; No. of Pages 7 2
ARTICLE IN PRESS M.C. Aznar et al. / European Journal of Radiology xxx (2014) xxx–xxx
been proposed as an alternative to PET/CT [3,4], whereby the increased soft tissue contrast through the MR and the overall reduction in patient exposure, primarily from the lack of ionizing radiation exposure from the CT, are considered key drivers for this technology [5]. Today, three different design concepts exist for PET/MRI [6–8] of which the fully-integrated concept [6] offers the ability to acquire PET and MRI data simultaneously, in addition to the advantages described above. Following the introduction of the first commercial PET/MRI system in 2010, clinical validation of these systems is still ongoing [5], and several technical challenges mandate further investigation [5,9]. One of the key challenges of combined PET/MRI, next to the modifications of the system hardware, is the lack of a CT-like transmission source and the subsequent need to perform attenuation and scatter correction using the available MR and emission information only. In the co-planar design, for example, MR-based attenuation correction (MR-AC) is performed using a dedicated T1weighted MR sequence in combination with a customized image processing technique to derive attenuation maps for whole-body PET [7]. The resulting MR-based attenuation map is composed of three classes of voxels: air (outside the subject), lungs and soft tissue. In the integrated design MR-AC is performed using an in-/opposed-phase MR sequence followed by an image-based segmentation [10] into four tissue classes: air (outside the subject), lungs, soft tissue and fat. Of note, this approach does not account for the presence of bone in the attenuation map. The absence of bone, however, could lead to a noticeable overall and local underestimation of the attenuation correction factors and, therefore, to a bias of the AC-PET data. Martinez-Möller et al. [11] simulated the absence of bone attenuation using PET/CT data from 35 oncology patients and concluded that the bias of AC-PET, measured as relative changes in standardized uptake values (SUV), was highest for lesions in the pelvic bone with an average SUV underestimation of 13%. The mean SUV for bone lesions across all patients decreased by 8% when using the segmented attenuation map without accounting for bone. Samarin and colleagues studied the effect of ignoring bone attenuation during MR-AC further [12]. They report an average decrease in SUV in bony structures and soft tissue lesions adjacent to bone of 11% (range: 2% to 31%) and 3% (0% to 4%), respectively relative to CT-based AC. The authors also demonstrate a dependency of the SUV bias on the composition of the bone lesions. Despite a general underestimation of the SUV in combined whole-body PET/MRI images none of the studies above reported on missing out on bone lesions entirely from standard MR-based AC [13]. Nonetheless, full quantitative accuracy is required for reliable therapy response assessment and further work to ensure proper representation of the bones is warranted. New developments suggest that bone segmentation based solely on MR images should be possible, at least for the brain region when employing UTE sequences [14] or atlas-based approaches [15]. However, in either case, assigning a relevant and accurate density distribution to the segmented bone structures is challenging. The goal of this study was to assess the effect of MR-AC on the accuracy and visibility of PET positive lesions in PET images of oncology patients undergoing integrated PET/MRI imaging. Here, we assume the availability of MR image volumes that represent particular bone structures as acquired with hypothetical MR sequences. Corresponding MR-based attenuation maps are derived from standard MR-AC maps following the 4-class tissue mode (background, lung tissue, fat, soft tissue) [11] with several segmented bone attenuation maps superimposed from separate CT scans, thereby assuming that such bone maps will become available through novel MR acquisition and processing in the future. We assessed SUV in reference tissues and lesions of PET/MRI images of oncology patients with lesions close to or within bony structures
following MR-AC using the original MR-based attenuation map and the modified attenuation maps. 2. Materials and methods 2.1. Patients Twenty clinical oncology patients with a referral for a PET/CT study were included in this study. Of those 10 patients with soft tissue lesions in the head/neck region were scheduled for an [18F]FDG examination (Group A) and 10 patients with osseous lesions were scheduled for a whole-body [18F]-NaF PET/CT examination (Group B). All patients underwent combined PET/CT and PET/MRI imaging between February 2012 and February 2013. 2.2. Imaging protocol The study protocol corresponded to a single-injection, dualimaging protocol and was approved by the local ethics committee. All patients underwent PET/CT imaging (Biograph TruePoint/mCT, Siemens) first. PET/CT imaging was performed without time-offlight (TOF) and images were reconstructed without using spatial resolution recovery [16], as this option was not available on subsequent PET/MRI. Patients in Group A were injected with 306 MBq ± 80 MBq of [18F]-FDG. Patients were positioned with their arms down and fixed with a thermoplastic mask over the head/neck region. PET/CT imaging started at 62 min ± 4 min post-injection. First, a topogram scan was acquired for the delineation of the co-axial imaging range, followed by a spiral CT scan (120 kV p, approximately 200 effective mAs adjusted after Siemens CareDose) and a 2-bed emission acquisition (2–3 min/bed). Patients in Group B were injected with 203 MBq ± 21 MBq of [18F]-NaF. Patients were positioned with their arms down. PET/CT imaging started at 49 min ± 5 min post-injection. First, a topogram scan was acquired for the delineation of the co-axial imaging range, followed by a spiral CT scan (120 kV p, 40 mA s) and a multi-bed (four to seven bed positions) emission acquisition (2 min/bed). Following the PET/CT examination patients were repositioned on the integrated PET/MRI system (Biograph mMR, Siemens). All patients were positioned head-first supine and with their arms down. In head/neck patients the thermoplastic mask was not used but the neck was supported by a radiotherapy cushion to provide similar positioning of the head/neck region, while allowing the use of the head/neck coil for MR. PET/MRI imaging started at 117 min ± 27 min post-injection for 20 min per bed position for patients in Group A and 130 min ± 32 min post-injection for 5 min per bed position for patients in Group B. The clinical imaging protocol included a scout scan for defining the co-axial imaging range, in-phase and opposed-phase MR imaging, using the Dixon VIBE sequence, for deriving MR-based attenuation factors (orgMR-AC) [11] as well as simultaneously acquired emission data and T1/T2 MR acquisitions. Protocol details are given in [17]. For the purpose of this study, PET data from the PET/MRI acquisition were corrected for attenuation using the orgMR-AC method as well as attenuation maps derived from co-registered CT data from the prior PET/CT examination. 2.3. Co-registration and bone segmentation All co-registration and segmentation procedures were performed in MATLAB (version 2011b, MathWorks, USA) and DICOM files were converted using the CERR package [18]. The CT images from the PET/CT examination were co-registered to the MR attenuation map (orgMR-AC) using a rigid transformation followed by a non-rigid transformation (cubic b-spline). All
Please cite this article in press as: Aznar MC, et al. Whole-body PET/MRI: The effect of bone attenuation during MR-based attenuation correction in oncology imaging. Eur J Radiol (2014), http://dx.doi.org/10.1016/j.ejrad.2014.03.022
G Model EURR-6724; No. of Pages 7
ARTICLE IN PRESS M.C. Aznar et al. / European Journal of Radiology xxx (2014) xxx–xxx
3
Fig. 1. Derivation of the attenuation maps with different representations of bone. Bone values were derived from the available CT image data: original HU values for bone (“orgCT”), average of the original HU values (“meanCT”) of the segmented CT volume, fixed value of 350 HU (“spongCT”) and fixed value of 1200 HU (“cortCT”). orgMC-AC refers to the original in-/opposed-phase attenuation map (no bones present).
co-registration steps were performed using MATLAB using a 4level normalized cross-correlation and a regularization parameter of 10−8 [19]. After the co-registration step, bone structures were extracted from the available CT images using a k-means clustering algorithm, which clustered the image intensities into four classes (air, soft tissue, fat and bone) and defined a centroid value representation the mean intensity of the bones. This centroid value was then used as input for a Potts model to assign each voxel to a specific class by including information about neighbouring voxels. As a last step, a morphological closing operation was performed to remove potential gaps and holes within the segmented skeleton.
2.4. CT-derived attenuation maps Four different strategies were investigated regarding the representation of bone attenuation values and subsequent attenuation correction (Fig. 1). The first attenuation map (orgCT) was based on the co-registered original CT attenuation values (HU). For the second attenuation map (meanCT) all values in the segmented bone contour were assigned the mean CT attenuation (HU) of the original bone voxels, simulating the attribution of bone density after a template-based segmentation [15]. Mean CT attenuation values were calculated per patient. The third and fourth attenuation maps employed the coregistered CT with the bone tissue values being replaced by fixed
Fig. 2. Definition of regions and volumes of interest for SUVmean: (A) for head and neck patients (Group A) and (B) for bone lesion patients (Group B). Reference regions and PET-positive lesions are shown in blue and red, respectively. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)
Please cite this article in press as: Aznar MC, et al. Whole-body PET/MRI: The effect of bone attenuation during MR-based attenuation correction in oncology imaging. Eur J Radiol (2014), http://dx.doi.org/10.1016/j.ejrad.2014.03.022
G Model EURR-6724; No. of Pages 7 4
ARTICLE IN PRESS M.C. Aznar et al. / European Journal of Radiology xxx (2014) xxx–xxx
Fig. 3. Case study for head and neck patients (Group A). Top row: series of attenuation maps according to Fig. 1 (converted to HU); bottom row: AC-PET images (in counts). Note, markedly different uptake in the bone region but little relative differences in the soft tissue lesion (arrow).
CT attenuation values for spongiosa and cortical bone, respectively, and simulating the attribution of bone density following a UTEbased segmentation [14]. The attenuation values for spongious and cortical bone were set to 350 HU and 1200 HU independent of the patient [20]. The resulting attenuation maps were named spongCT and cortCT (Fig. 1). Each of the three alternative approaches to MR-AC (meanCT, spongCT and cortCT) corresponds to a clinical situation where bone structures would be classified on dedicated MR acquisitions, e.g., [14], while ignoring intra-tissue variations of physical bone density. Attenuation values at 511 keV were calculated for each voxel in the four attenuation maps above by applying the segmentation, bi-linear scaling approach by Carney et al. [21]. The modified maps (i.e., including the segmented bones from CT) were inserted in the original MR-AC DICOM files replacing the original values. The modified DICOM files were then imported into the reconstruction computer according to instructions provided by the vendor. PET/MRI images were reconstructed following AC using orgMRAC and any of the four derived attenuation maps above. Image reconstruction was based on 3D-OSEM with 3 iterations and 21 subsets on 344 by 344 matrices with a 4 mm Gaussian filter applied.
2.5. Image analysis We assessed the relative changes of SUV in reference regions and lesions of PET/MRI images following AC using any of the three modified attenuation maps meanCT, spongCT and cortCT as well as the PET images following AC using standard MR-AC (orgMRAC) and the reference orgCT. Fig. 2 summarizes the location of the regions (ROI and VOI) used for the region-based analysis of PET images of patients in Group A and B. All regions were defined manually by an experienced nuclear medicine physician on the PET/MRI images following orgMR-AC using the Mirada software (version XD3.4, Mirada Medical Ltd., UK). In Group A (Fig. 2A), reference regions were defined in the upper spine and in the cerebellum using an ROI in the axial plane and a spherical VOI, respectively. Uptake within the lesion was calculated from an isocontour at 40% of SUVmax. In the bone lesions patients (Group B), reference regions (see Fig. 2B) were defined as the bladder (spherical VOI in the centre of the organ), the femoral head (spherical VOI), the spine (ROI in the axial plane) and the pelvic bone (ROI in the coronal plane). Similar to Group A, uptake within the lesion was calculated from an isocontour at 40% of SUVmax.
Fig. 4. Relative difference (%) of average SUVmean in reference regions and lesions for head and neck patients in Group A. SB—spinal bone.
Please cite this article in press as: Aznar MC, et al. Whole-body PET/MRI: The effect of bone attenuation during MR-based attenuation correction in oncology imaging. Eur J Radiol (2014), http://dx.doi.org/10.1016/j.ejrad.2014.03.022
G Model EURR-6724; No. of Pages 7
ARTICLE IN PRESS
19.8 (6.8; 63.2) 1.3 (−9.0; 13.5) 15.0 (5.8; 55.4) 17.7 (6.8; 57.6) 12.5 (5.4; 52.7) 13.9 (5.7; 57.7)
−10.0 (−22.4; 0.0)
16.5 (−4.5; 55.3)
RD (%) cortCT
12.6 (2.5; 26.0) −2.2 (−13.7; 0.0)
RD (%) spongCT
10.8 (2.2; 23.0) −7.2 (−22.5; 0.0)
RD (%) meanCT
12.8 (2.4; 25.3)
RD (%)
9.9 (2.2; 21.8)
orgMR-AC
Patients in Group A included 3 women and 7 men, with an average age of 62 years (range 53;72) and an average weight of 76 kg (range 38;123). Fig. 3 shows a case study of a head/neck patient from Group A processed with the four different attenuation maps (Fig. 1). The PET-positive lesion was visible across the different PET images. Table 1 summarizes the relative differences in SUVmean for all soft tissue lesions in patients from Group A with respect to values from orgCT. In all but one patient, orgMR-AC caused an underestimation of the SUVmean value with respect to orgCT (see Table S1, Supplementary material). In contrast, the other CTbased AC method tended to lead to an overestimation of SUVmean values. There was a large variation of apparent lesion uptake across patients. The differences in SUVmean between orgMR-AC and orgCT ranged from 0.0% to −22.5% (median −7.2%). The maximum median difference (12.6%) was observed between orgCT and cortCT though meanCT could result in as high a bias as cortCT in individual patients (see Table 1). Fig. 4 summarizes the relative changes in SUVmean averaged across all head/neck patients in Group A for the reference regions and soft tissue lesions. Relative differences were smaller in the cerebellum, on average −14% (range −17% to −8%) for SUVmean between orgMR-AC and orgCT. The upper spine is the region most affected by variations in AC with a relative difference in SUVmean between orgMR-AC and orgCT of −27% across patients (range: −40% to −8%). This relative difference was as high as 165% for cortCT in the upper spine. Patients in Group B included 4 women and 6 men, with an average age of 62 years (range 34;75) and an average weight of 84 kg (range 68;110). Fig. 5 shows a case study of a patient with bone lesions from Group B processed with the four attenuation maps (Fig. 1). The PET-positive lesion was seen on all PET images but exaggerated on cortCT and spongCT. Table 1 summarizes also the relative differences in SUVmean for the bone lesions in each patient of Group B with respect to values from orgCT. Relative differences of SUVmean of the lesions were strongly affected by the different AC maps: the maximum relative difference ranges from −22.4% with orgMR-AC to 110.8% with cortCT. On average, values obtained with spongCT displayed the smallest bias with respect to orgCT (1.3%). Fig. 6 summarizes the relative differences in SUVmean of reference regions and lesions in patients in Group B. The largest difference with respect to orgCT was observed for cortCT for all regions. The bladder was visibly unaffected by the differences in the AC maps, with an average bias of 2% for SUVmean between orgMR-AC and orgCT (the maximum bias observed in an
orgCT
3. Results
11.3 (2.2; 23.5)
with SUVrec being the SUVmean for a given attenuation map (orgMR-AC, meanCT, spongCT or cortCT)) and SUVorgCT being the corresponding SUVmean in PET images following orgCT, considered to be the gold standard. All SUV were related to the patient body weight. Relative differences were reported for each reference region and lesion. Here, reference region referred to an area of physiologic tracer uptake in an organ (e.g., bladder) or in a substructure of an organ (e.g., bone) as illustrated in Fig. 2. We calculated the mean relative difference (%) as the average across all 10 patients in Group A and all 10 patients in Group B for reference regions, bone lesions and soft tissue lesions.
Head and neck lesions median (range) Bone lesions median (range)
(1)
SUVmean
SUVrec − SUVorgCT SUVorgCT
Table 1 SUVmean values of head and neck lesions (10 patients, Group A) and bone lesions (10 patients, Group B). Results are displayed as median and range. All relative differences are reported with respect to orgCT.
RD% = 100 ×
10.8 (1.6; 17.3)
For each of the four attenuation maps, relative differences (%) of attenuation-corrected PET images were assessed for all reference regions and lesions:
40.8 (1.5; 110.8)
5
13.4 (−2.3; 17.3)
M.C. Aznar et al. / European Journal of Radiology xxx (2014) xxx–xxx
Please cite this article in press as: Aznar MC, et al. Whole-body PET/MRI: The effect of bone attenuation during MR-based attenuation correction in oncology imaging. Eur J Radiol (2014), http://dx.doi.org/10.1016/j.ejrad.2014.03.022
G Model EURR-6724; No. of Pages 7
ARTICLE IN PRESS M.C. Aznar et al. / European Journal of Radiology xxx (2014) xxx–xxx
6
Fig. 5. Case study for bone lesion patients (Group B). Top row: series of attenuation maps according to Fig. 1 (converted to HU); bottom row: AC-PET images (in counts). Case includes a bone lesion in the left pelvis (arrow).
individual patient: 20% with cortCT) while the femoral head bone was the reference region most affected.
4. Discussion In this study, we assessed the effect of using different attenuation maps on PET image quality and accuracy of PET/MRI data of oncology patients. Our first objective was similar to that of previous studies by Samarin [12] and Martinez-Möller [11]. However, here we use data from fully-integrated PET/MRI. Issues such as potential PET/MR-specific artifacts and correction for the presence of MR coils in the PET attenuation do not mandate special attention as they are independent of the relative measures (Eq (1)) derived from alterations of the patient-based attenuation maps.
In our patient cohort, we observed a similar bias in soft tissue lesions close to the bone and in bone lesions of −7.2% and −10% (range 0%; −22%). Here, we assumed orgCT to be the gold standard. While this value supports previous reports for the bias in bone lesions [11,12], it is slightly higher than those reported by Samarin [12] for soft tissue lesions. In their study they reported a bias of −3.2% (range 0%; −4%), though the histology in their patient group was not specified. In our cohort, all soft tissue lesions were squamous cell carcinoma of the head and neck. Large tumours and smaller positive lymph nodes were present in this cohort and one cannot exclude that a potential bias as high as 22% might cause a small positive lesion to be undetected. Moreover, uptake in reference bone regions was also affected by not accounting for bone during MR-AC. In our patient cohort, this bias was most pronounced in the upper spine of patients in Group
Fig. 6. Relative difference (%) of average SUVmean in reference regions and lesions for head and neck patients in Group B. Femoral HB—femoral head bone.
Please cite this article in press as: Aznar MC, et al. Whole-body PET/MRI: The effect of bone attenuation during MR-based attenuation correction in oncology imaging. Eur J Radiol (2014), http://dx.doi.org/10.1016/j.ejrad.2014.03.022
G Model EURR-6724; No. of Pages 7
ARTICLE IN PRESS M.C. Aznar et al. / European Journal of Radiology xxx (2014) xxx–xxx
A. In contrast the cerebellum and especially the bladder were relatively unaffected in patients in Group A and Group B, respectively. Our second objective was to investigate which density to assign to the segmented bone structures using MR-attenuation maps corresponding to variations of existing as well as hypothetical approaches to MR-AC. For example, orgCT, would correspond to an attenuation map derived from MR with the full anatomical variability of attenuation values of soft tissue and bone for which no standard MR-only method is available today. The attenuation data in spongCT and cortCT would represent soft tissue values as derived from in/opposed phase imaging merged with a segmented bone representation (e.g., from an UTE sequence) whereby all bone values are set to fixed attenuation values representing spongiosa bone or cortical bone. Using the information from a separate CT for segmentation is a time-consuming procedure and is not an option for routine clinical adoption but reconstructing PET images with simple in/opposed phase imaging (orgMR-AC) already results in a bias in both bone lesion and soft tissue lesion uptake. Our results suggest that assigning a uniform density corresponding to spongious bone is a reasonable compromise. Indeed, in bone lesions, the overestimation of SUVmean with respect to orgCT was highest for cortCT (40.8%, range: 1.5% to 110.8%) while spongCT values were the closest to orgCT (average bias of 1.3%, range: −9.0 to 13.5%). For soft tissue lesions the bias was highest with cortCT (13.4%, range: −2–17%) but lowest with spongCT (−2−2%, range 0.0% to −13.7%). In average across all patients, the bias in SUVmean between spongCT and orgCT was less than 5% for lesions in Group A and in Group B. This strategy could provide a realistic alternative to a separate CT acquisition assuming that the bones can be accurately segmented. A limitation of our study is that our procedure is sensitive to co-registration and segmentation errors, mainly due to positioning differences between the CT and the PET/MRI examination. However, on visual examination, patient position was very reproducible between PET/CT and PET/MR and, thus, residual misalignment was not considered to be a major source of error. Our results suggest that an accurate representation of the bones is needed to assure the quantitative accuracy of PET/MRI data. In the absence of bone attenuation factors, the accuracy of the quantification of bone lesions, as well as soft tissue lesions relatively close to the bone, is inaccurate. If bone segmentation as obtained via a template-based approach or from dedicated UTE sequences is available, then assigning a fixed attenuation value of spongious bone to the whole skeleton results in only a minor bias (
