Home
Search
Collections
Journals
About
Contact us
My IOPscience
Lean body mass correction of standardized uptake value in simultaneous whole-body positron emission tomography and magnetic resonance imaging
This content has been downloaded from IOPscience. Please scroll down to see the full text. 2015 Phys. Med. Biol. 60 4651 (http://iopscience.iop.org/0031-9155/60/12/4651) View the table of contents for this issue, or go to the journal homepage for more
Download details: IP Address: 157.89.65.129 This content was downloaded on 30/05/2015 at 22:38
Please note that terms and conditions apply.
Institute of Physics and Engineering in Medicine Phys. Med. Biol. 60 (2015) 4651–4664
Physics in Medicine & Biology doi:10.1088/0031-9155/60/12/4651
Lean body mass correction of standardized uptake value in simultaneous whole-body positron emission tomography and magnetic resonance imaging Thies H Jochimsen1, Jessica Schulz1,2, Harald Busse3, Peter Werner1, Alexander Schaudinn3, Vilia Zeisig1, Lars Kurch1, Anita Seese1, Henryk Barthel1, Bernhard Sattler1 and Osama Sabri1 1
Department of Nuclear Medicine, Leipzig University Hospital Liebigstr. 18, D-04103 Leipzig, Germany 2 Max Planck Institute for Human Cognitive and Brain Sciences Stephanstr. 1a, D-04103 Leipzig, Germany 3 Department of Diagnostic and Interventional Radiology, Leipzig University Hospital Liebigstr. 20, D-04103 Leipzig, Germany E-mail: [email protected] Received 24 June 2014, revised 8 April 2015 Accepted for publication 21 April 2015 Published 28 May 2015 Abstract
This study explores the possibility of using simultaneous positron emission tomography—magnetic resonance imaging (PET-MRI) to estimate the lean body mass (LBM) in order to obtain a standardized uptake value (SUV) which is less dependent on the patients' adiposity. This approach is compared to (1) the commonly-used method based on a predictive equation for LBM, and (2) to using an LBM derived from PET-CT data. It is hypothesized that an MRIbased correction of SUV provides a robust method due to the high soft-tissue contrast of MRI. A straightforward approach to calculate an MRI-derived LBM is presented. It is based on the fat and water images computed from the two-point Dixon MRI primarily used for attenuation correction in PET-MRI. From these images, a water fraction was obtained for each voxel. Averaging over the whole body yielded the weight-normalized LBM. Performance of the new approach in terms of reducing variations of 18F-Fludeoxyglucose SUVs in brain and liver across 19 subjects was compared with results using predictive methods and PET-CT data to estimate the LBM. The MRI-based method reduced the coefficient of variation of SUVs in the brain by 41 ± 10% which is comparable to the reduction by the PET-CT 0031-9155/15/124651+14$33.00 © 2015 Institute of Physics and Engineering in Medicine Printed in the UK
4651
T H Jochimsen et al
Phys. Med. Biol. 60 (2015) 4651
method (35 ± 10%). The reduction of the predictive LBM method was 29 ± 8%. In the liver, the reduction was less clear, presumably due to other sources of variation. In conclusion, employing the Dixon data in simultaneous PET-MRI for calculation of lean body mass provides a brain SUV which is less dependent on patient adiposity. The reduced dependency is comparable to that obtained by CT and predictive equations. Therefore, it is more comparable across patients. The technique does not impose an overhead in measurement time and is straightforward to implement. Keywords: PET-MRI, SUV, lean body mass, Dixon sequence (Some figures may appear in colour only in the online journal) Introduction In positron emission tomography (PET), the standardized uptake value (SUV) is used to quantify the uptake of a radioactive tracer in tissue (Keyes 1995). It is usually defined as the measured radioactive concentration in tissue divided by the injected activity per total body weight (BW). As the uptake of many radiotracers, such as 18F-Fludeoxyglucose (FDG), is generally small in adipose tissue, more reliable SUVs might be obtained by considering only the lean body mass (LBM), i.e. the total body weight minus the fat weight, instead of the whole-body mass. It has already been shown that the variation of FDG SUVs is reduced when calculated in this way (Zasadny and Wahl 1993). Usually, the LBM is derived using a simple empiric formula (Sugawara et al 1999), but this method can only give a rough estimate of the fraction of non-fat tissue. As predictive equations may cause substantial errors at an individual level(see Results below for an example) (Erselcan et al 2002), a better way would be to actually measure the LBM for each patient. In recent studies by Chan (2012), Hamill et al (2013) and Kim et al (2013), the non-fat tissue content of the body was determined from anatomical data acquired with a combined PET—computed tomography (PET-CT) scanner. Fat tissue could be discriminated from other tissues based on different x-rays attenuation coefficients. This study explores the possibility to use simultaneous PET—magnetic resonance imaging (PET-MRI) to measure the LBM. It is hypothesized that an MRI-based correction of SUV provides a robust method due to the high soft-tissue contrast of MRI. Preliminary results of this study were previously presented at a conference (Sattler et al 2013). Adipose tissue can be segmented from MRI data in several ways: For instance, T1-weighted imaging provides a very high contrast between fat and other soft tissue because of the large underlying differences in relaxation times (Brennan et al 2005). Alternatively, the Dixon method is very specific as it is based on the differences in chemical shift of the fat and water magnetic resonance signal (Dixon 1984). In its simplest form (two-point Dixon), it consists of two MRI acquisitions at two different echo times (TEs). The TEs are adjusted so as to yield two images: In the first, the magnetizations from fat and water protons are in phase, i.e. they add up. In the second, the signals are of opposite phase, i.e. they are subtracted. This set of two signal equations can be solved voxel-wise for the fat and water signal, providing fat and water images. Practically, a more sophisticated scheme has to be used to account for magnetic field inhomogeneities (Coombs et al 1997). Using data from the Dixon sequence for correction of SUV has several advantages over other methods in simultaneous whole-body PET-MRI: In contrast to methods where voxels are explicitly assigned to a certain tissue type (Brennan et al 2005, Chan 2012, Hamill et al 2013), tissues/organs which contain both fat and water in 4652
T H Jochimsen et al
Phys. Med. Biol. 60 (2015) 4651
Figure 1. Schematic description of the method to calculate the MRI-based SUV according to equation (5). Coronal slices of one patient are shown. The operator mean(...)Mask denotes calculating the mean value over all voxels contained in the mask. Please note how the water and fat images are influenced by the inhomogeneous sensitivity profiles of the receiver coils, while the normalized map of water fraction is not.
a similar range, as for instance the liver (visible in figure 1), contribute partially to the LBM. Furthermore, the MRI-based LBM correction of the SUV can be conducted without additional CT radiation exposure and with little or no increase of the total scan time as the Dixon method is currently used for attenuation correction of the PET data (Martinez-Möller et al 2009). In this work, a straightforward approach to calculate an MRI-based LBM is presented. Its performance, in terms of reducing variations of brain FDG SUV, is compared with the other methods mentioned before. In this context we would like to point out that, although PET-MRI and PET-CT data of the same subjects is presented, it is not the purpose of this study to compare PET-MRI SUVs with PET-CT SUVs. Rather, the inclusion of PET-CT data was motivated by the possibility to compare different methods for LBM-correction of SUVs, and to apply the correction to SUVs of both modalities. Methods All PET-MRI scans analyzed in the present study were acquired within clinical routine during whole-body imaging sessions carried out in oncology patients. As approved by the local ethics committee, patients were asked to undergo an additional PET-MRI examination after a routine PET-CT scan. Prior to examination, patients had given informed consent that their data will be used for research purposes. Simultaneous whole-body PET-MRI data were acquired using a clinically-certified (FDA approval and CE mark) 3-Tesla Biograph mMR (Siemens Healthcare, Erlangen, 4653
T H Jochimsen et al
Phys. Med. Biol. 60 (2015) 4651
Germany). The PET component has a field-of-view (FOV) of 258 mm (axial) × 594 mm (transaxial). Up to six bed positions were acquired, separated by 197 mm (30.5 mm overlap). In each bed position, the PET data were acquired over 5 min. These data were reconstructed into a 256 × 256 matrix (voxel size: 2.32 × 2.32 × 2.03 mm3) using the built-in 3D ordered subset expectation maximization (OSEM) algorithm with 3 iterations, 21 subsets and a 3 mm Gaussian filter. During PET, a two-point MRI Dixon sequence with a FOV of 500 mm (L–R), 333 mm (H–F) and 333 mm (A–P) and an isotropic voxel size of 2.6 mm was acquired. For attenuation correction of the PET data during reconstruction, attenuation coefficient maps (air, lung, soft tissue, fat) were segmented from the fat and water images generated by the Dixon sequence. Other anatomical MRI sequences for T1-, T2- and diffusion-weighted MRI followed in each bed position. The reconstructed image data were composed to whole-body datasets by the system. These datasets were used in the subsequent processing steps. Depending on patient position and examined regions, a total axial FOV of 98–145 cm, starting from the patients vertex of the head, was covered by the Dixon sequence in this way. Any artifacts in the fat/water images of the Dixon sequence will not only influence the calculated MRI-based LBM in a direct way, but also the SUV indirectly via an erroneous attenuation correction of the PET data. The latter effect could bias the results when comparing different methods to calculate the SUV. Therefore, as a control experiment, the present study used PET data of the same subject from a PET-CT examination (utilizing a standard CT-based attenuation correction) acquired on the same day using the same injection of radiotracer. In the PET-CT examination, the PET data were acquired over 3–4 min in each bed position and reconstructed into a 168 × 168 matrix (voxel size: 4.07 × 4.07 × 3.0 mm3) using 3D OSEM with 4 iterations, 8 subsets and a 5 mm Gaussian filter. The transaxial FOV was 683 mm. PET-MRI and PET-CT data from 19 patients (10 male, 9 female, 31–75 years) were collected for this study. As FDG brain SUVs were used for comparison (see below), patients with major neurological deficits were prospectively excluded from the study. Another two patients were excluded retrospectively because they underwent chemotherapy during or three months before the examination leading to uncertainty on possible chemotherapychanges of FDG delivery to the brain via altered hemodynamics/excretion of the tracer. Moreover, several patients with obvious segmentation errors in the attenuation maps from the Dixon sequence, which occurred frequently during the first months of operating the PET-MRI and would result in severe PET quantification errors, were either not added to the study in the first place or excluded retrospectively. A dose of 258–441 MBq FDG was injected intravenously (mean and standard deviation over subjects: 364 ± 62 MBq). The uptake time (duration from injection to start of scan) was 119 ± 38 min for PET-CT and 180 ± 53 min for PET-MRI. The SUV based on total body weight, SUVBW, is defined as C SUV (1) BW = A /BW
with C as the tissue activity concentration and A as the amount of injected activity. Assuming a tissue density of 1.0 g ml−1, SUVs are dimensionless throughout this paper. SUVLBM denotes the SUV based on the predicted lean body mass (Sugawara et al 1999), LBMpred, and is calculated as C SUV = wpred ⋅ SUVBW LBM = (2) A /LBMpred 4654
T H Jochimsen et al
Phys. Med. Biol. 60 (2015) 4651
with wpred = LBMpred/BW as the relative LBM. LBMpred is usually calculated from the patients weight (BW in kg), size (H in cm) and gender using the formula (Chan 2012, Sugawara et al 1999) 2 ⎧ for male, ⎪ 1.1 ⋅ BW − 128 ⋅ (BW/H) , ⎨ LBM = (3) pred ⎪ ⎩1.07 ⋅ BW − 148 ⋅ (BW/H)2, for female.
Similarly, The CT-based SUV is defined by SUV (4) CT = wCT ⋅ SUVBW
with wCT as the relative CT-based LBM which will be defined below in equation (6). It is analogous to wpred in equation (2). Finally, the SUV based on the MRI-derived LBM, SUVMRI, is defined accordingly: SUV (5) MRI = w MRI ⋅ SUVBW
Here, wMRI is the MRI-based water fraction. The procedure to derive wMRI from the MRI data is described in the following. A strategy to calculate wMRI from Dixon-based fat and water images would be to compare the total signal intensities of the fat and the water images. However, due to inhomogeneities in the sensitivities of the radio-frequency receiving coils in a whole-body examination, regions covered with high sensitivity and a predominant tissue composition (e.g. a sensitive head coil in combination with mostly water in the brain, as in figure 1) would bias the result. Therefore, wMRI was calculated in the following way, as illustrated in figure 1: first, the fat and water images were added together. This sum corresponds to the field-inhomogeneity-corrected inphase Dixon data. The resulting image was used to automatically create a mask containing all tissue voxels. For that, the threshold of the image intensity to outline the mask was determined from a histogram of image intensities (100 bins). The threshold was set to the first minimum above zero in this histogram so that the ‘noise floor’ of the image is excluded, as illustrated in figure 2. This step is necessary since MRI signal intensities do not relate directly to a physical quantity to which a simple fixed threshold could be applied. Although this method is not universal, the resulting masks gave good outlines of the tissue regions, as verified by visual inspection. Then, the water fraction of each voxel was calculated by dividing the value of the water image by the sum of water and fat image. In this way, the calculated value is not influenced by the strongly inhomogeneous sensitivity profiles of the receiver coils (as visible in figure 1). Finally, wMRI was calculated by averaging over all voxels of the mask. To compare the SUVs across subjects, the brain was chosen as the primary target organ for several reasons: Firstly, it intrinsically exhibits a high FDG uptake, i.e. it provides a high signal-to-noise (SNR) ratio due to a high PET count rate. Secondly, it is a relatively large organ which is less prone to major partial-volume and spill-over effects. Thirdly, it is a rigid structure which can be easily registered from one modality to the other, and tools for doing so are readily available, as described in the following. Lastly, attenuation due to the head coil is compensated for in the PET reconstruction by a template attenuation map. In contrast, flexible surface coils used in other parts of the body are not accounted for and are known to lead to errors in PET quantification. If not mentioned otherwise, the Object-oriented Development Interface for NMR (Jochimsen and von Mengershausen 2004) was used for all data processing steps. The brain’s voxels were extracted by applying the Brain Extraction Tool of the Functional MRI of the Brain Software Library (FSL) (Smith 2002) to the top of the Dixon water images. The resulting masks were used in the subsequent processing steps. Maps of partial volume coefficients 4655
T H Jochimsen et al
Phys. Med. Biol. 60 (2015) 4651
1e+07
Threshold for Mask
1e+06
Voxel Count
1e+05 10000 1000 100 10 1
0
20
40 60 Bin of Signal Intensity
80
100
Figure 2. Example histogram of whole-body image intensities to generate the mask in figure 1. The hatched bars contain the voxels of the noise floor which are excluded for the mask. Please note the logarithmic scale of the y-axis.
of gray matter (GM), white matter (WM) and cerebrospinal fluid were generated from the Dixon water images by using the FSL Automated Segmentation Tool (Zhang et al 2001). These maps were later used to calculate the mean SUVs corrected for partial-volume effects in whole brain (GM and WM combined) by inverting the geometric transformation matrix (Rousset et al 1998) using the known point-spread function of the PET subsystem (Delso et al 2011). Maps of SUVBW were calculated from the PET-MRI and PET-CT data using PMOD 3.2 (PMOD Technologies, Zurich, Switzerland). The maps from PET-CT were registered to the maps from PET-MRI using the FSL Linear Image Registration Tool (Jenkinson and Smith 2001). Finally, mean SUVs in whole brain were calculated using the procedure described above. As a result, mean SUVs, corrected for partial volume effects, are obtained for each modality. For comparison, a method to obtain LBM from the CT data, similar to that by Chan (2012) and Hamill et al (2013), was applied as follows: The axially reconstructed CT data of the PET-CT examination were analyzed by a custom-made volumetry software developed under IDL (Version 6.2, Exelis VIS, Boulder, CO). This tool automatically displays the CT histogram of a graphically defined region of interest (ROI) and allows for an interactive definition (both input fields and graphical) of individual Hounsfield ranges. The corresponding voxels are simultaneously overlaid on the original CT slice for visual inspection by a radiological expert. As coverage of the legs by CT varied, the axial region from head to pubic symphysis was used in the analysis of all patients. In a first step, regions like the CT table were excluded from analysis. The previously reported Hounsfield unit (HU) ranges (Chan 2012) for compartments involving air (−1000 – −191 HU), fat tissue (−190 – −30 HU), soft tissue (−29 – +151 HU) and bones (+152– + 2500 HU) generated overlay regions that agreed well with visual impression and were therefore used for further analysis. Corresponding tissue volumes were automatically computed by the number of pixels in the respective histogram range multiplied by the pixel dimensions (area) and distance between slice centers (here, equal to slice thickness). Average tissue densities of 0.26, 0.89, 1.02 and 1.12 g mL−1 for lung, fat, soft tissue and bones, respectively, were assumed to obtain the tissue masses (Saito et al 2012). This required an additional, crude manual segmentation of all axial slices involving the lung. Finally, the CT-based weight-normalized LBM (wCT in equation (4)) was estimated by 4656
T H Jochimsen et al
Phys. Med. Biol. 60 (2015) 4651
Mlung + Msoft + Mbone w CT = LBM CT/BW = (6) Mlung + Msoft + Mbone + Mfat
where the overall tissue masses Mlung, Msoft, Mbone, and Mfat of lung tissue, soft tissue, bone and fat, respectively, were obtained by multiplying the number of pixels with the corresponding tissue densities. While the CT-based SUV correction was derived from the LBM of only the head and the trunk, the MRI-derived SUV correction takes also part of the legs into account. To make both comparable, the same axial region (from head to pubic symphysis) as in the CT-based method was extracted from the MRI Dixon data and the MRI-based SUV correction was repeated using this cropped data. To test the method in another organ than the brain, mean and maximum SUVBW were obtained from physiological liver tissue as described before (Heusch et al 2013). Briefly, using dedicated viewing software (Syngo.Via, Siemens Healthcare, Erlangen, Germany) 3 cm3 circular volumes of interest (VOIs) were placed manually in segment 7 of the liver. In 2 of 19 patients, activity in the liver had to be quantified in another segment due to focal liver pathology in segment 7. The VOIs were placed in PET-CT and PET-MRI by 4 experienced readers. As outliers were relatively frequent (i.e. the result from one reader deviated decisively from those of the other three), the median SUVBW were calculated across readers, instead of the less robust mean. Finally, the same SUV correction methods as for the brain were applied. To compare the methods quantitatively in terms of their performance to reduce the intersubject variations in SUVs, it is important to note that the correction not only alters the variance, but also the mean of the SUVs. Hence, it is necessary to use the coefficient of variation (COV), i.e. the standard deviation as a fraction of the mean over subjects, when comparing the performance. Because a statistical analysis based on error propagation is not possible due to large number of unknown sources of variation, the standard error (deviation) of the COV was estimated by statistics applied to jackknife resampling, i.e. by recalculating the COV after SUV correction for each subsample which leaves out one patient at a time. Using this standard deviation, a one-tailed, one-sample t-test was used to test for a significant (p
Search
Collections
Journals
About
Contact us
My IOPscience
Lean body mass correction of standardized uptake value in simultaneous whole-body positron emission tomography and magnetic resonance imaging
This content has been downloaded from IOPscience. Please scroll down to see the full text. 2015 Phys. Med. Biol. 60 4651 (http://iopscience.iop.org/0031-9155/60/12/4651) View the table of contents for this issue, or go to the journal homepage for more
Download details: IP Address: 157.89.65.129 This content was downloaded on 30/05/2015 at 22:38
Please note that terms and conditions apply.
Institute of Physics and Engineering in Medicine Phys. Med. Biol. 60 (2015) 4651–4664
Physics in Medicine & Biology doi:10.1088/0031-9155/60/12/4651
Lean body mass correction of standardized uptake value in simultaneous whole-body positron emission tomography and magnetic resonance imaging Thies H Jochimsen1, Jessica Schulz1,2, Harald Busse3, Peter Werner1, Alexander Schaudinn3, Vilia Zeisig1, Lars Kurch1, Anita Seese1, Henryk Barthel1, Bernhard Sattler1 and Osama Sabri1 1
Department of Nuclear Medicine, Leipzig University Hospital Liebigstr. 18, D-04103 Leipzig, Germany 2 Max Planck Institute for Human Cognitive and Brain Sciences Stephanstr. 1a, D-04103 Leipzig, Germany 3 Department of Diagnostic and Interventional Radiology, Leipzig University Hospital Liebigstr. 20, D-04103 Leipzig, Germany E-mail: [email protected] Received 24 June 2014, revised 8 April 2015 Accepted for publication 21 April 2015 Published 28 May 2015 Abstract
This study explores the possibility of using simultaneous positron emission tomography—magnetic resonance imaging (PET-MRI) to estimate the lean body mass (LBM) in order to obtain a standardized uptake value (SUV) which is less dependent on the patients' adiposity. This approach is compared to (1) the commonly-used method based on a predictive equation for LBM, and (2) to using an LBM derived from PET-CT data. It is hypothesized that an MRIbased correction of SUV provides a robust method due to the high soft-tissue contrast of MRI. A straightforward approach to calculate an MRI-derived LBM is presented. It is based on the fat and water images computed from the two-point Dixon MRI primarily used for attenuation correction in PET-MRI. From these images, a water fraction was obtained for each voxel. Averaging over the whole body yielded the weight-normalized LBM. Performance of the new approach in terms of reducing variations of 18F-Fludeoxyglucose SUVs in brain and liver across 19 subjects was compared with results using predictive methods and PET-CT data to estimate the LBM. The MRI-based method reduced the coefficient of variation of SUVs in the brain by 41 ± 10% which is comparable to the reduction by the PET-CT 0031-9155/15/124651+14$33.00 © 2015 Institute of Physics and Engineering in Medicine Printed in the UK
4651
T H Jochimsen et al
Phys. Med. Biol. 60 (2015) 4651
method (35 ± 10%). The reduction of the predictive LBM method was 29 ± 8%. In the liver, the reduction was less clear, presumably due to other sources of variation. In conclusion, employing the Dixon data in simultaneous PET-MRI for calculation of lean body mass provides a brain SUV which is less dependent on patient adiposity. The reduced dependency is comparable to that obtained by CT and predictive equations. Therefore, it is more comparable across patients. The technique does not impose an overhead in measurement time and is straightforward to implement. Keywords: PET-MRI, SUV, lean body mass, Dixon sequence (Some figures may appear in colour only in the online journal) Introduction In positron emission tomography (PET), the standardized uptake value (SUV) is used to quantify the uptake of a radioactive tracer in tissue (Keyes 1995). It is usually defined as the measured radioactive concentration in tissue divided by the injected activity per total body weight (BW). As the uptake of many radiotracers, such as 18F-Fludeoxyglucose (FDG), is generally small in adipose tissue, more reliable SUVs might be obtained by considering only the lean body mass (LBM), i.e. the total body weight minus the fat weight, instead of the whole-body mass. It has already been shown that the variation of FDG SUVs is reduced when calculated in this way (Zasadny and Wahl 1993). Usually, the LBM is derived using a simple empiric formula (Sugawara et al 1999), but this method can only give a rough estimate of the fraction of non-fat tissue. As predictive equations may cause substantial errors at an individual level(see Results below for an example) (Erselcan et al 2002), a better way would be to actually measure the LBM for each patient. In recent studies by Chan (2012), Hamill et al (2013) and Kim et al (2013), the non-fat tissue content of the body was determined from anatomical data acquired with a combined PET—computed tomography (PET-CT) scanner. Fat tissue could be discriminated from other tissues based on different x-rays attenuation coefficients. This study explores the possibility to use simultaneous PET—magnetic resonance imaging (PET-MRI) to measure the LBM. It is hypothesized that an MRI-based correction of SUV provides a robust method due to the high soft-tissue contrast of MRI. Preliminary results of this study were previously presented at a conference (Sattler et al 2013). Adipose tissue can be segmented from MRI data in several ways: For instance, T1-weighted imaging provides a very high contrast between fat and other soft tissue because of the large underlying differences in relaxation times (Brennan et al 2005). Alternatively, the Dixon method is very specific as it is based on the differences in chemical shift of the fat and water magnetic resonance signal (Dixon 1984). In its simplest form (two-point Dixon), it consists of two MRI acquisitions at two different echo times (TEs). The TEs are adjusted so as to yield two images: In the first, the magnetizations from fat and water protons are in phase, i.e. they add up. In the second, the signals are of opposite phase, i.e. they are subtracted. This set of two signal equations can be solved voxel-wise for the fat and water signal, providing fat and water images. Practically, a more sophisticated scheme has to be used to account for magnetic field inhomogeneities (Coombs et al 1997). Using data from the Dixon sequence for correction of SUV has several advantages over other methods in simultaneous whole-body PET-MRI: In contrast to methods where voxels are explicitly assigned to a certain tissue type (Brennan et al 2005, Chan 2012, Hamill et al 2013), tissues/organs which contain both fat and water in 4652
T H Jochimsen et al
Phys. Med. Biol. 60 (2015) 4651
Figure 1. Schematic description of the method to calculate the MRI-based SUV according to equation (5). Coronal slices of one patient are shown. The operator mean(...)Mask denotes calculating the mean value over all voxels contained in the mask. Please note how the water and fat images are influenced by the inhomogeneous sensitivity profiles of the receiver coils, while the normalized map of water fraction is not.
a similar range, as for instance the liver (visible in figure 1), contribute partially to the LBM. Furthermore, the MRI-based LBM correction of the SUV can be conducted without additional CT radiation exposure and with little or no increase of the total scan time as the Dixon method is currently used for attenuation correction of the PET data (Martinez-Möller et al 2009). In this work, a straightforward approach to calculate an MRI-based LBM is presented. Its performance, in terms of reducing variations of brain FDG SUV, is compared with the other methods mentioned before. In this context we would like to point out that, although PET-MRI and PET-CT data of the same subjects is presented, it is not the purpose of this study to compare PET-MRI SUVs with PET-CT SUVs. Rather, the inclusion of PET-CT data was motivated by the possibility to compare different methods for LBM-correction of SUVs, and to apply the correction to SUVs of both modalities. Methods All PET-MRI scans analyzed in the present study were acquired within clinical routine during whole-body imaging sessions carried out in oncology patients. As approved by the local ethics committee, patients were asked to undergo an additional PET-MRI examination after a routine PET-CT scan. Prior to examination, patients had given informed consent that their data will be used for research purposes. Simultaneous whole-body PET-MRI data were acquired using a clinically-certified (FDA approval and CE mark) 3-Tesla Biograph mMR (Siemens Healthcare, Erlangen, 4653
T H Jochimsen et al
Phys. Med. Biol. 60 (2015) 4651
Germany). The PET component has a field-of-view (FOV) of 258 mm (axial) × 594 mm (transaxial). Up to six bed positions were acquired, separated by 197 mm (30.5 mm overlap). In each bed position, the PET data were acquired over 5 min. These data were reconstructed into a 256 × 256 matrix (voxel size: 2.32 × 2.32 × 2.03 mm3) using the built-in 3D ordered subset expectation maximization (OSEM) algorithm with 3 iterations, 21 subsets and a 3 mm Gaussian filter. During PET, a two-point MRI Dixon sequence with a FOV of 500 mm (L–R), 333 mm (H–F) and 333 mm (A–P) and an isotropic voxel size of 2.6 mm was acquired. For attenuation correction of the PET data during reconstruction, attenuation coefficient maps (air, lung, soft tissue, fat) were segmented from the fat and water images generated by the Dixon sequence. Other anatomical MRI sequences for T1-, T2- and diffusion-weighted MRI followed in each bed position. The reconstructed image data were composed to whole-body datasets by the system. These datasets were used in the subsequent processing steps. Depending on patient position and examined regions, a total axial FOV of 98–145 cm, starting from the patients vertex of the head, was covered by the Dixon sequence in this way. Any artifacts in the fat/water images of the Dixon sequence will not only influence the calculated MRI-based LBM in a direct way, but also the SUV indirectly via an erroneous attenuation correction of the PET data. The latter effect could bias the results when comparing different methods to calculate the SUV. Therefore, as a control experiment, the present study used PET data of the same subject from a PET-CT examination (utilizing a standard CT-based attenuation correction) acquired on the same day using the same injection of radiotracer. In the PET-CT examination, the PET data were acquired over 3–4 min in each bed position and reconstructed into a 168 × 168 matrix (voxel size: 4.07 × 4.07 × 3.0 mm3) using 3D OSEM with 4 iterations, 8 subsets and a 5 mm Gaussian filter. The transaxial FOV was 683 mm. PET-MRI and PET-CT data from 19 patients (10 male, 9 female, 31–75 years) were collected for this study. As FDG brain SUVs were used for comparison (see below), patients with major neurological deficits were prospectively excluded from the study. Another two patients were excluded retrospectively because they underwent chemotherapy during or three months before the examination leading to uncertainty on possible chemotherapychanges of FDG delivery to the brain via altered hemodynamics/excretion of the tracer. Moreover, several patients with obvious segmentation errors in the attenuation maps from the Dixon sequence, which occurred frequently during the first months of operating the PET-MRI and would result in severe PET quantification errors, were either not added to the study in the first place or excluded retrospectively. A dose of 258–441 MBq FDG was injected intravenously (mean and standard deviation over subjects: 364 ± 62 MBq). The uptake time (duration from injection to start of scan) was 119 ± 38 min for PET-CT and 180 ± 53 min for PET-MRI. The SUV based on total body weight, SUVBW, is defined as C SUV (1) BW = A /BW
with C as the tissue activity concentration and A as the amount of injected activity. Assuming a tissue density of 1.0 g ml−1, SUVs are dimensionless throughout this paper. SUVLBM denotes the SUV based on the predicted lean body mass (Sugawara et al 1999), LBMpred, and is calculated as C SUV = wpred ⋅ SUVBW LBM = (2) A /LBMpred 4654
T H Jochimsen et al
Phys. Med. Biol. 60 (2015) 4651
with wpred = LBMpred/BW as the relative LBM. LBMpred is usually calculated from the patients weight (BW in kg), size (H in cm) and gender using the formula (Chan 2012, Sugawara et al 1999) 2 ⎧ for male, ⎪ 1.1 ⋅ BW − 128 ⋅ (BW/H) , ⎨ LBM = (3) pred ⎪ ⎩1.07 ⋅ BW − 148 ⋅ (BW/H)2, for female.
Similarly, The CT-based SUV is defined by SUV (4) CT = wCT ⋅ SUVBW
with wCT as the relative CT-based LBM which will be defined below in equation (6). It is analogous to wpred in equation (2). Finally, the SUV based on the MRI-derived LBM, SUVMRI, is defined accordingly: SUV (5) MRI = w MRI ⋅ SUVBW
Here, wMRI is the MRI-based water fraction. The procedure to derive wMRI from the MRI data is described in the following. A strategy to calculate wMRI from Dixon-based fat and water images would be to compare the total signal intensities of the fat and the water images. However, due to inhomogeneities in the sensitivities of the radio-frequency receiving coils in a whole-body examination, regions covered with high sensitivity and a predominant tissue composition (e.g. a sensitive head coil in combination with mostly water in the brain, as in figure 1) would bias the result. Therefore, wMRI was calculated in the following way, as illustrated in figure 1: first, the fat and water images were added together. This sum corresponds to the field-inhomogeneity-corrected inphase Dixon data. The resulting image was used to automatically create a mask containing all tissue voxels. For that, the threshold of the image intensity to outline the mask was determined from a histogram of image intensities (100 bins). The threshold was set to the first minimum above zero in this histogram so that the ‘noise floor’ of the image is excluded, as illustrated in figure 2. This step is necessary since MRI signal intensities do not relate directly to a physical quantity to which a simple fixed threshold could be applied. Although this method is not universal, the resulting masks gave good outlines of the tissue regions, as verified by visual inspection. Then, the water fraction of each voxel was calculated by dividing the value of the water image by the sum of water and fat image. In this way, the calculated value is not influenced by the strongly inhomogeneous sensitivity profiles of the receiver coils (as visible in figure 1). Finally, wMRI was calculated by averaging over all voxels of the mask. To compare the SUVs across subjects, the brain was chosen as the primary target organ for several reasons: Firstly, it intrinsically exhibits a high FDG uptake, i.e. it provides a high signal-to-noise (SNR) ratio due to a high PET count rate. Secondly, it is a relatively large organ which is less prone to major partial-volume and spill-over effects. Thirdly, it is a rigid structure which can be easily registered from one modality to the other, and tools for doing so are readily available, as described in the following. Lastly, attenuation due to the head coil is compensated for in the PET reconstruction by a template attenuation map. In contrast, flexible surface coils used in other parts of the body are not accounted for and are known to lead to errors in PET quantification. If not mentioned otherwise, the Object-oriented Development Interface for NMR (Jochimsen and von Mengershausen 2004) was used for all data processing steps. The brain’s voxels were extracted by applying the Brain Extraction Tool of the Functional MRI of the Brain Software Library (FSL) (Smith 2002) to the top of the Dixon water images. The resulting masks were used in the subsequent processing steps. Maps of partial volume coefficients 4655
T H Jochimsen et al
Phys. Med. Biol. 60 (2015) 4651
1e+07
Threshold for Mask
1e+06
Voxel Count
1e+05 10000 1000 100 10 1
0
20
40 60 Bin of Signal Intensity
80
100
Figure 2. Example histogram of whole-body image intensities to generate the mask in figure 1. The hatched bars contain the voxels of the noise floor which are excluded for the mask. Please note the logarithmic scale of the y-axis.
of gray matter (GM), white matter (WM) and cerebrospinal fluid were generated from the Dixon water images by using the FSL Automated Segmentation Tool (Zhang et al 2001). These maps were later used to calculate the mean SUVs corrected for partial-volume effects in whole brain (GM and WM combined) by inverting the geometric transformation matrix (Rousset et al 1998) using the known point-spread function of the PET subsystem (Delso et al 2011). Maps of SUVBW were calculated from the PET-MRI and PET-CT data using PMOD 3.2 (PMOD Technologies, Zurich, Switzerland). The maps from PET-CT were registered to the maps from PET-MRI using the FSL Linear Image Registration Tool (Jenkinson and Smith 2001). Finally, mean SUVs in whole brain were calculated using the procedure described above. As a result, mean SUVs, corrected for partial volume effects, are obtained for each modality. For comparison, a method to obtain LBM from the CT data, similar to that by Chan (2012) and Hamill et al (2013), was applied as follows: The axially reconstructed CT data of the PET-CT examination were analyzed by a custom-made volumetry software developed under IDL (Version 6.2, Exelis VIS, Boulder, CO). This tool automatically displays the CT histogram of a graphically defined region of interest (ROI) and allows for an interactive definition (both input fields and graphical) of individual Hounsfield ranges. The corresponding voxels are simultaneously overlaid on the original CT slice for visual inspection by a radiological expert. As coverage of the legs by CT varied, the axial region from head to pubic symphysis was used in the analysis of all patients. In a first step, regions like the CT table were excluded from analysis. The previously reported Hounsfield unit (HU) ranges (Chan 2012) for compartments involving air (−1000 – −191 HU), fat tissue (−190 – −30 HU), soft tissue (−29 – +151 HU) and bones (+152– + 2500 HU) generated overlay regions that agreed well with visual impression and were therefore used for further analysis. Corresponding tissue volumes were automatically computed by the number of pixels in the respective histogram range multiplied by the pixel dimensions (area) and distance between slice centers (here, equal to slice thickness). Average tissue densities of 0.26, 0.89, 1.02 and 1.12 g mL−1 for lung, fat, soft tissue and bones, respectively, were assumed to obtain the tissue masses (Saito et al 2012). This required an additional, crude manual segmentation of all axial slices involving the lung. Finally, the CT-based weight-normalized LBM (wCT in equation (4)) was estimated by 4656
T H Jochimsen et al
Phys. Med. Biol. 60 (2015) 4651
Mlung + Msoft + Mbone w CT = LBM CT/BW = (6) Mlung + Msoft + Mbone + Mfat
where the overall tissue masses Mlung, Msoft, Mbone, and Mfat of lung tissue, soft tissue, bone and fat, respectively, were obtained by multiplying the number of pixels with the corresponding tissue densities. While the CT-based SUV correction was derived from the LBM of only the head and the trunk, the MRI-derived SUV correction takes also part of the legs into account. To make both comparable, the same axial region (from head to pubic symphysis) as in the CT-based method was extracted from the MRI Dixon data and the MRI-based SUV correction was repeated using this cropped data. To test the method in another organ than the brain, mean and maximum SUVBW were obtained from physiological liver tissue as described before (Heusch et al 2013). Briefly, using dedicated viewing software (Syngo.Via, Siemens Healthcare, Erlangen, Germany) 3 cm3 circular volumes of interest (VOIs) were placed manually in segment 7 of the liver. In 2 of 19 patients, activity in the liver had to be quantified in another segment due to focal liver pathology in segment 7. The VOIs were placed in PET-CT and PET-MRI by 4 experienced readers. As outliers were relatively frequent (i.e. the result from one reader deviated decisively from those of the other three), the median SUVBW were calculated across readers, instead of the less robust mean. Finally, the same SUV correction methods as for the brain were applied. To compare the methods quantitatively in terms of their performance to reduce the intersubject variations in SUVs, it is important to note that the correction not only alters the variance, but also the mean of the SUVs. Hence, it is necessary to use the coefficient of variation (COV), i.e. the standard deviation as a fraction of the mean over subjects, when comparing the performance. Because a statistical analysis based on error propagation is not possible due to large number of unknown sources of variation, the standard error (deviation) of the COV was estimated by statistics applied to jackknife resampling, i.e. by recalculating the COV after SUV correction for each subsample which leaves out one patient at a time. Using this standard deviation, a one-tailed, one-sample t-test was used to test for a significant (p
