Improving the CT (140 kVp) to PET (511 keV) conversion in PET/MR Hardware Component Attenuation Correction.

PURPOSE
Today, attenuation correction (AC) of PET/MR hardware components is performed by using an established method from PET/CT hybrid imaging. As shown in previous studies, the established mathematical conversion from CT to PET attenuation coefficients may, however, lead to incorrect results in PET quantification when applied to AC of hardware components in PET/MR. The purpose of this study is to systematically investigate the attenuating properties of various materials and electronic components frequently used in the context of PET/MR hybrid imaging. The study, thus, aims at improving hardware component attenuation correction in PET/MR.


MATERIAL AND METHODS
Overall, 38 different material samples were collected; a modular phantom was used to for CT, PET, and PET/MR scanning of all samples. CT-scans were acquired with a tube voltage of 140 kVp to determine Hounsfield Units (HU). PET transmission scans were performed with 511 keV to determine linear attenuation coefficients (LAC) of all materials. The attenuation coefficients were plotted to obtain an HU to LAC correlation graph, which was then compared to two established conversions from literature. Hardware attenuation maps of the different materials were created and applied to PET data reconstruction following a phantom validation experiment. From these measurements, PET difference maps were calculated to validate and compare all three conversion methods.


RESULTS
For each material, the HU and corresponding LAC could be determined and a bi-linear HU to LAC conversion graph was derived. The corresponding equation was y = 1.64 * 10-5 × (HU + 1000) + 8.3 * 10-2 . While the two established conversions lead to a mean quantification PET bias of 4.69 % ± 0.27 % and -2.84 % ± 0.72 % in a phantom experiment, PET difference measurements revealed only 0.5 % bias in PET quantification when applying the new conversion resulting from this study.


CONCLUSION
An optimized method for the conversion of CT to PET attenuation coefficients has been derived by systematic measurement of 38 different materials. In contrast to established methods, the new conversion also considers highly attenuating materials, thus improving attenuation correction of hardware components in PET/MR hybrid imaging.


INTRODUCTION
Since its introduction in 2010, simultaneous positron emission tomography/magnetic resonance (PET/MR) has found its way into various clinical applications ranging from neuro to pediatric and cardiac to whole-body applications in oncology. [1][2][3][4] Recent studies also work towards the integration of PET/MR hybrid imaging into the concept of radiation therapy planning. 5,6 All of these PET/MR applications have in common, that for high-quality PET and MR imaging dedicated hardware is required: radiofrequency (RF) coils, phantoms, positioning aids, and table platforms. [7][8][9][10][11][12] During simultaneous PET and MR data acquisition, these hardware components are positioned in the field-of-view (FOV) of the PET detector and attenuate the PET annihilation photons. This may lead to visible artifacts and bias in PET quantification. 11 Consequently, all PET signal attenuation caused by hardware components and patient tissue needs to be corrected using appropriate attenuation correction (AC) methods. These AC methods in PET/MR are fundamentally different compared to PET/CT hybrid imaging. Attenuation correction in PET/MR is generally performed with two different approaches. Photon attenuation due to human tissues is based on MR imaging using dedicated AC sequences that provide a segmentation into different tissue classes and then assign linear attenuation coefficients (LAC) for each tissue class. [13][14][15][16] Photon attenuation due to hardware components, on the other hand, is based on computed tomography (CT) scans of the respective hardware component. 17,18 The CT data is then converted with dedicated conversion curves to the energy level of the PET component. 19 Both methods for AC in PET/MR are thus different from AC in PET/CT. Computed tomography can provide high-resolution transmission scans. In PET/CT these CT scans represent the spatial distribution of attenuating human tissues and hardware components such as the PET/CT patient table, phantoms, and further table platforms. Furthermore, PET signal attenuating RF coils are not needed for signal reception in PET/CT. Today, the AC of human tissues in PET/MR is usually solved by applying dedicated MR sequences, subsequent tissue segmentation integrating bone models, atlas-based methods and the emission-based estimation of attenuation. [13][14][15][16]20,21 Moreover, joint reconstruction algorithms add to the palette of methods for human tissue AC, and deep learning now enters the field to provide MR-derived pseudo-CT data for AC of human tissues. 22,23 However, AC of hardware components such as RF coils, phantoms, and radiation therapy (RT) equipment has to be performed differently, since all these components and materials are not visible in MR imaging and thus cannot be segmented and assigned with attenuation coefficients. The established approach in PET/MR is to acquire CT-based three-dimensional (3D) virtual attenuation models of the respective hardware component (e.g., RF coil). 24,25 These CT-based AC templates represent the spatial distribution of attenuation values in Hounsfield units (HU), acquired at a specific CT tube voltage of, for example, 140 kVp. The attenuation values of the CTbased AC template are then converted to linear attenuation coefficients (LAC) at 511 keV PET energy levels. This method for hardware AC in PET/MR resembles the established method for attenuation correction in PET/CT.
Numerous mathematical HU to LAC conversion methods have been suggested for application in PET/CT, including bilinear conversion curves, quadratic polynomial calibration curve, dual-energy CT and conversions based on virtual dualenergy CT data (VDECT). [26][27][28][29] All these methods aim at accurate conversion of the measured CT data, for example, acquired at 80-140 kVp, to proper linear attenuation coefficients (LAC) at 511 keV energy level of PET data. Accurate conversion, thus, is a precondition to obtaining accurate quantification of PET data following AC. The bi-linear conversion methods proposed by Kinahan et al. and Carney et al. today are routinely established in PET/CT hybrid imaging and work fast, accurate and robust for AC of patient tissues and phantom measurements in PET/CT. 17,18 While the established conversions work well in PET/CT, all these methods only focused on the conversion of lower attenuation values of human tissues and are not optimized for the correction of higher attenuating materials and metals. 18,19,[26][27][28][29] With the advent of the PET/MR systems, however, additional hardware components such as the patient table and radiofrequency coils and further materials may present higher attenuation values than human tissues. Thus, Paulus et al., in their study extended the HU scale during CT acquisition towards higher attenuating materials (up to 30 000 HU). 19 The study by Paulus et al. demonstrated, that considering also the stronger attenuating hardware components and materials results in an adapted slope of the curve provided by Carney et al. and therefore reduces bias in CTbased attenuation maps of hardware components for use in PET/MR applications. 18,19 Building on the findings by Paulus et al., this study aims to provide an optimized conversion curve for AC of ancillary hardware components in PET/MR hybrid imaging. Varying materials ranging from plastics and metals, as well as electronic components and cables, were systematically investigated. The use of these diverse materials can all be found in the manufacture of RF coils, phantoms, and other ancillary hardware. The attenuating properties of 38 different material probes were investigated in both 140 kVp CT scans and 511 keV PET transmission scans. A new HU-to-LAC-conversion curve was derived and compared to the established conversion curves. 18,19 The results of this study lead to an adapted and improved HU-to-LAC conversion curve that also considers highly attenuating materials and, consequently, further improves the accuracy of AC of the hardware setup in PET/MR hybrid imaging.

2.A. Material samples
A total of 38 material samples were examined. All material samples were provided by a RF coil manufacturer (Rapid Biomedical GmbH, Rimpar, Germany) and include various plastics, metals, cables, and electronic components commonly used in PET/MR phantom housing and RF coil manufacturing. The material samples were divided into two major groups of 19 samples each (Fig. 1). The first group comprises heterogeneous samples (#1-#19), mainly electronic components, including ceramic capacitors of different sizes, cables, soldering tin, inductors, small RF preamplifiers, and others. The second group includes solid and homogeneous samples (#20-#38) like various plastics and metals, such as aluminum, copper, and brass, used for hardware component manufacturing (Fig. 1).
The group of heterogeneous samples includes electronic components that all vary in size and shape. For better comparison between the two groups, thin-walled plastic test tubes with a cylindrical shape and comparable dimensions (16 mm outer diameter, length 200 mm) like the homogeneous samples were selected. The electronic components were filled Medical Physics, 47 (5), May 2020 into the test tubes and sealed with flexible foam, to compress and fixate the small movable electronic components minimizing displacement between experiments.
The homogenous samples (#20-#38) are each 150 mm long and have a diameter of 16 mm. The unified sample size and cylindrical shape were considered as advantageous regarding handling in phantom experiments: cylindrical shape for uniform photon transmission, and symmetry for designing a modular phantom setup. The entire list of materials and detailed specifications is provided in Table I.

2.B. Phantom design
A phantom was designed that allows for reproducible transmission measurements of all 38 material samples in CT and PET systems. Polystyrene (PS) foam was selected so it would not significantly influence the CT and PET transmission measurements due to its mechanically stable but rather low-attenuating properties. The average HU for the PS material was determined to be À900, which is close to the HU of air (HU À1000). An annular-shaped holder was chosen to permit the equidistant and symmetric arrangement of the samples around the isocenter axis of the respective CT, PET, and PET/MR systems during all measurements. Through the process of water jet cutting, the PS foam was formed into an annular shape with an inner diameter of 214 mm and an outer diameter of 294 mm. The ring is supported by a little pedestal at the base providing stability during CT, PET, and PET/ MR measurements. Two identical annular phantom holders were manufactured and 19 equidistant holes were drilled into the PS foam to provide a reproducible setting for the 19 heterogeneous An inner diameter of 214 mm was chosen for the purpose of allowing the cylindrically shaped water phantom (volume 9480 ml) to be used in conjunction with the overall phantom setting [ Fig. 2(c)]. Additionally, the water phantom can be filled with radiotracer through ports located on the top lid, thus providing a homogeneous large-volume emission source.

2.C. CT scans
The two sample holders loaded with all 38 materials were first scanned on a dual-energy CT scanner (SOMATOM Force, Siemens Healthcare GmbH, Erlangen, Germany). The dual-energy option of the CT system was neither needed nor used in this experiment. 140 kVp was chosen for CT energy, which is the highest energy level that is broadly available on all modern CT and PET/CT hybrid imaging systems. This assures the applicability of the results of this study at other institutions as well as comparability to the results of the studies by Carney et al. 21 and by Paulus et al. 22 The following CT imaging parameters were used: tube voltage 140 kVp, tube current 650 mA, matrix size of 512 9 512 pixels, voxel size 0.3 9 0.3 9 0.6 mm 3 and a B30f-smooth convolution kernel. Reconstruction of the CT data was done by applying the extended Hounsfield scale. While the standard CT HU scale ranges from À1024 HU to +3071 HU to include all human body, (background air up to the cortical bone, calcifications, and teeth), the extended CT scale displays HU in steps of ten up to a maximum of +30 710 HU. Therefore, most high attenuating materials and hardware components including different metals and ceramics are covered by the extended CT scale.
The CT data for the material samples were examined with VINCI (VINCI Version 4.63, Max-Planck-Institute for Metabolism Research, Cologne, Germany). 30 The aim was to identify the mean HU value for each material sample, scanned with 140 kVp tube voltage. Therefore, the Dicom CT data were overlaid with a region-of-interest (ROI) of 16 mm diameter in the axial view [ Fig. 3(a)]. The mean HU value for each ROIs, each with a slice thickness of 0.6 mm, was measured over the entire length of the material sample. All mean values of all ROI in each slice were taken into account for determining the overall mean HU value for the specific material [ Fig. 3(a)]. An additional threshold (À100 HU) for the heterogeneous materials was applied with MATLAB 2013b (MATrix LABoratory, MathWorks, MA, USA) using a customized MATLAB tool, to segment and consider only the full material components within the sample tubes and not surrounding air between the individual components.

2.D. PET transmission scans
A dedicated PET scanner, ECAT EXACT HR + , Siemens Healthcare GmbH, Erlangen, Germany) equipped with rotating 68 Ge/ 68 Ga rod sources (each around 150 MBq) was used to obtain LAC from all materials. 31 The two-dimensional linear attenuation coefficients of all materials are directly measured by the transmission of the rotating line sources while the PET detector was operated in transmission mode. 32 No further attenuation correction or energy conversion had to be applied because the measured attenuation coefficients are the reference LAC values acquired at 511 keV.
For the PET transmission measurements two sample holders with the heterogeneous and homogeneous probes were fixed free-floating at the head end of the patient table and two PET measurements were performed sequentially, one for each sample holder. Acquisition time was more than 150 min to obtain a high number of counts and true events. The data were reconstructed applying OSEM (10 iterations, 8 subsets, Gauss filter 6.0 mm) using the e7-tools (Siemens Molecular Imaging, Knoxville, USA) with a matrix of 128 x 128 and 63 slices, resulting in a pixel size of 5.2 x 5.2 mm 2 and a slice thickness of 2.43 mm.
The 3D PET data were spatially coregistered and matched with the VINCI software onto the 3D CT data to obtain exact spatial co-registration of PET and CT data for subsequent accurate conversion of HU and LAC values. The high spatial resolution of the CT data serves as a positioning reference for the placement of the ROI to obtain the mean LAC values in the PET data. Figure 3 depicts the coregistration and positioning of ROIs in CT and PET data along two imaging planes to obtain volumes of interest (VOI) in the material samples (Fig. 3). Thus, LAC values were measured over a volume of interest to average spatial inhomogeneities in the heterogeneous samples.

2.E. Bilinear conversion
Equation (1) The conversion, from the measured HU to LAC using the two bilinear conversions [Eqs. (1) and (2) (MathWorks, Natick, USA). The following post-processing steps were performed as described in more detail in Paulus et al. work. 19 To reduce the potential quantitative impact of minor streaking artifacts between neighboring samples on CT measurements (HU), for each sample the average HU value of surrounding artifacts was determined and subtracted from the measured HU for each sample. The masking of each circular sample was performed by thresholding with À600 HU and placing a circular mask on each sample. All residual background voxels exceeding the actual geometry of each sample were set to zero. Finally, a 4 mm Gauss filter was applied to all CT data for smoothing and matching the high spatial resolution of the CT data to the lower spatial resolution PET data. 19 The two conversion methods according to Eqs. (1) and (2)  Additionally, the measured HU values and measured corresponding LAC for each of the 38 samples are plotted in a HU to LAC-conversion graph to result in a curve reflecting the CT and PET measurements of this study. By connecting the measured values with lines of best fit, an adapted equation for the HU to LAC conversion can be derived for the results. This can then be compared to the reference curves suggested by Carney et al. and by Paulus et al. To finally validate the measured HU to LAC conversion and to compare it to the two reference conversion methods, all three conversions are used to generate µmaps from CT data for both the heterogeneous and homogeneous material samples. To match the high spatial resolution of the CT data with the lower spatial resolution of the PET detector (4.0-6.0 mm), a 4.0 mm Gaussian filter was applied during µmap generation. The three µmaps are then applied for attenuation correction of the material samples in PET/MR difference measurements.

2.F. PET/MR difference measurements
To evaluate and validate the effectiveness of the three different HU to LAC conversion methods, PET difference measurements were performed on an integrated 3-Tesla PET/MR whole-body hybrid imaging system (Biograph mMR, Siemens Healthcare GmbH, Erlangen, Germany). For this purpose, the cylindrical water phantom was filled with 18-F isotope to serve as a PET emission source. The phantom was positioned in the isocenter of the PET detector without any attenuation interference from the PET/MR patient table by mounting it free-floating at the head-end of the PET/MR patient table [ Fig. 4(a)].
The phantom activity (230 MBq in total, concentration 24 kBq/ml) and the extended acquisition time of 35 min. in the phantom experiments were chosen to provide sufficient PET signal statistics. The setup without material samples served as the reference scan without any attenuation. The two sample holders were then consecutively placed onto the emission phantom [ Fig. 4(b)]. Two PET emission measurements were carried out, one for each sample holder, to determine the 511 keV attenuation properties of each material sample. Both measurements were then time-and decay-corrected, to equalize the impact of tracer decay over time. Thus, in the PET difference measurements, a PET emission measurement without attenuating samples served as a reference; two further PET emission measurements with attenuating samples surrounding the water emission phantom provide data to determine the attenuation at 511 keV of each sample. For subsequent evaluation of the three different HU to LAC conversion curves, the CT-based 3D attenuation maps of the heterogeneous and homogeneous phantom samples are applied for attenuation correction of the PET emission data of both phantoms. In the ideal situation, an optimal conversion would provide absolute correction for the attenuation from all materials and would result in a homogeneous PET signal across the entire large volume water PET emission phantom. PET difference measurements between perfect attenuation corrected PET data with samples and the PET reference scan acquired without attenuating samples would, thus, result in zero difference across the water phantom.
The PET data were acquired in list mode. The reconstructed imaging matrix size was 344 x 344 x 127 with a voxel size of 2.09 x 2.09 x 2.03 mm 3 . The reconstruction parameters of 3 subsets and 21 iterations correspond to the 3D ordinary Poisson ordered-subsets expectation maximization (3D OP-OSEM) reconstruction parameters.
Furthermore, the CT-based 3D template model of the cylindrical water phantom was used for attenuation correction of the phantom. The six AC PET data sets (PET material ) (two material groups, three µmaps for each material group) were then compared to the reference PET scan (PET reference ), where only the cylindrical water phantom was attenuation corrected, by calculating difference maps [Eq. (3)].
All post-processing steps and data reconstructions of the PET data were performed offline using the e7 tools software provided by the manufacturer of the PET/MR hybrid system. due to the high spatial resolution of the CT scan. The homogeneous materials (lower row) can be depicted as regular shaped cylindrical structures with a smooth surface and homogeneous cross-sections (Fig. 5).

3.B. PET transmission scans
The two consecutive 511 keV PET transmission scans of the 2x 19 material samples provide linear attenuation coefficients of all material samples. In Fig. 5 all heterogeneous materials (upper row) can be seen in the axial view of PET transmission images (right). The lowest measured LAC is PUR SG95 (polyurethane plates for RF coil housing manufacturing, sample #19) with 0.082 cm À1 AE 0.005 cm À1 LAC, whereas the highest attenuating heterogeneous material is soldering tin with 0.296 cm À1 AE 0.006 cm À1 LAC (sample #4). Figure 5 depicts the axial view from PET transmission scans of the 19 homogenous material samples (lower row), where the high attenuating rods can be identified as copper and brass (samples #22 and #31). The measured LAC range from polyamide with a LAC of 0.0724 cm À1 AE 0.0034 cm À1 LAC (sample #33) to copper with 0.56 cm À1 AE 0.015 cm À1 LAC (sample #22).
The HU values for the different materials, given in Table II, are measured mean values with their respective standard deviation.

3.C. Bilinear conversion
The measured LAC values of the material samples were matched to the corresponding HU values. Figure 6 shows a HU to LAC-conversion graph with three data sets, which are all based on the measured HU values of all 38 material samples. The use of the Carney conversion method provides the highest slope in the HU to LAC conversion graph (blue crosses). The data points with the lowest slope (red circles) are Paulus converted LAC from measured CT-data HU. The third set of data points (green triangles) illustrates the acquired CT HU and the LAC from the PET transmission scans as obtained. The HU and LAC values were connected by a line of best fit to derive a corresponding conversion curve from the measured data points.
For the line of best fit of measured HU and LAC values, the slope and y-intercept are calculated and shown in Eq.  Only the right slope of the bi-linear curve was derived and adapted from the results of the present study. Equation (4) provides the line of best fit for the measured LAC values depicted in Fig. 6.

3.D. PET/MR difference measurements
The results of PET/MR phantom difference measurements are represented in Fig. 7. The relative difference maps show the calculated difference [Eq. (3)] between the following two measurements: first, a measurement of the emission phantom, without any material samples (reference scan); and, second, a measurement with PET signal attenuating material samples placed around the emission phantom and with applied attenuation correction using the three different µmaps. The PET difference measurements with AC according to Carney et al. comes to a mean overall difference of 4.69% AE 0.27% as measured in a large circular ROI across the entire cross-section of the water phantom and measured in 60 slices along the z-axis of the phantom. Positive difference values and the associated red color in the phantom here indicate a resulting overcorrection when applying Carney AC. Applying Paulus AC results in a mean difference of 2.84% AE 0.72% as measured within the water emission phantom volume. This indicates an undercorrection for Paulus AC.
Finally, applying the attenuation correction µmap as derived from HU and LAC measurements in this work leads to a low mean difference of 0.45% AE 0.21%, reflecting lower deviations between measured and true LAC values than for the Carney and Paulus conversions.
The line graphs in Fig. 7 depict the variation of the mean difference measured in the circular ROI of each cross-sectional slice of the PET difference data sets along multiple slices over a distance of 120 mm along the z-axis of the phantom. These graphs show that the measured mean difference for the three conversions is relatively stable along the z-axis of the phantom.
The second row of Fig. 7 represents the respective mean difference values for the homogeneous material samples. Carney AC shows a mean difference of 1.15% AE 0.18%. Paulus AC gives a mean difference of À2.67% AE 0.25%. Finally, applying the Oehmigen AC µmap as derived from HU and LAC measurements resulted in a low mean difference of 0.36% AE 0.12%.

DISCUSSION
The current study systematically investigated both the CT and PET attenuation coefficients of a broad spectrum of materials covering a large range from low to highly attenuating materials. CT attenuation coefficients of highly attenuating materials were captured by using the extended HU-scale covering high values up to +30 000 HU. The overlay of HU and LAC attenuation coefficients derived from the conversions by Carney et al., by Paulus et al. together with the measured attenuation coefficients derived from this research work reveals that the slope of the conversion curve derived from this study is in between the two previously published methods. 18,19 At the same time, the new conversion derived from this study is closer to the conversion suggested by Paulus et al. than it is to the PET/CT conversion proposed earlier by Carney et al..

4.A. CT
As expected, the standard deviations (SD) of measured HU in the homogeneous samples were low, whereas the measured HU in the heterogeneous samples shows a remarkably higher SD. The small dimensions of the collection of multiple electronic components in the test tubes leads to local HU maxima and HU minima, which result in higher SD when compared to the homogenous material samples (Table II). As a limitation of the CT measurements, it has to be noted that the homogeneous metal samples with a diameter of 16 mm strongly attenuate the CT photons (copper with +24 000 HU and brass with +27 000 HU). In these situations, such high attenuation coefficients paired with a rather large sample diameter may hamper accurate CT value reconstruction and can introduce artifacts. Thus, deviations between measured and real attenuation coefficients of the highly attenuating metal samples cannot be excluded.

4.B. PET
PET difference measurements serve as a highly sensitive validation experiment for the measured HU, measured LAC, and derived conversion curve. A large-volume (9.5 L) active emission phantom serves as reference for the PET measurement. Using the same phantom and identical position, a second PET measurement was taken with the modular phantom ring which included 19 heterogeneous or 19 homogeneous material samples. This however, will show lower activity values in the PET emission phantom due to the photon attenuation by various material samples surrounding the phantom. Under ideal conditions, reconstruction of the second PET scan with an appropriate attenuation correction will show a 0% difference across the whole volume of the phantom between the two PET measurements. Creating µmaps according to the methods by Carney et al. and by Paulus et al., and subsequent use of the µmaps derived in this research lead to a visible and measurable PET quantification bias in the volume of the PET emission phantom (Fig. 7). The blue color in the PET difference maps indicates a general undercorrection, whereas the red color indicates the overcorrection of the true attenuation coefficients of the various material samples. The red and blue color is not evenly distributed across the phantom volume (Fig. 7). In general, the upper parts of the phantom show more blue color, while the lower parts of the phantom show more red color. This indicates that not all material samples are corrected according to their true attenuation coefficients when using the conversions by Carney et al. and by Paulus et al. (Fig. 7, first and second column). When applying the conversion curve derived in this study to attenuation correction of heterogeneous and homogeneous materials, the overall PET difference bias across the phantom is much lower, as indicated by less red and blue color and by increased homogeneity across the phantom volume (Fig. 7, third column). The last column represents the measured PET quantification bias across the phantom. As a result a large circular ROI was placed on the cross-sectional view of the emission phantom and the overall bias within the ROI was determined. This was repeated for multiple transaxial slices along the z-axis of the phantom. The ROI measurements revealed a mean difference of 4.69% reduced mean difference of 0.45% AE 0.21%, reflecting only low deviations. Due to these results, it should be emphasized that the large ROIs placed in the cross-sectional difference images of the emission phantom averages positive and negative bias. Local deviations close to a single material sample can be much higher (undercorrection or overcorrection). In clinical PET/MR applications, the resulting local bias in PET quantification can be much higher when, for example, SUV measurements are performed close to the attenuating structures of an RF coil that has not been corrected according to its "real" attenuation values. In this context, the faint red and blue color in the difference maps for the conversion derived indicates that each material sample (heterogeneous or homogeneous) was corrected with a derived attenuation value that is close to the true physical attenuation value.
Although the described AC methods provide correction factors for different materials and hardware components located within the FOV of the PET detector, the method of AC has general limitations when larger amounts of highly attenuating materials are to be corrected. Such samples irreversibly reduce the PET signal. Consequently, applying significant AC factors will amplify noise and introduce artifacts, which may hamper PET quantification in the resulting corrected images as when using the metal samples in Fig. 7. Thus, the primary aim of designing RF coils and other hardware components for PET/MR hybrid imaging is to avoid excessive photon in advance. 24,25 This is achieved by using appropriate hardware design which avoids the placement of larger amounts of highly attenuating materials (e.g. metals) within the PET field-of-view. Only once an optimized and PET-transparent RF coil design has been constructed, can the remaining attenuation be corrected according to the method and conversion curve suggested in this study.
Such an elaborate experimental study involving numerous steps does come with potential limitations in each of the experimental steps. The development and realization of an experiment for conversion curve adaption involved several steps like the selection of appropriate material samples, phantom design and the execution of all measurements. The involved imaging modalities CT, PET, and PET/MR all come with inherent physical characteristics that need to be considered during planning and execution of an experiment. CT and PET data post-processing includes mathematical conversions, image co-registration, artifact removal, thresholding, and filtering. Furthermore, evaluation of the resulting CT and PET image data included ROI and VOI definition, averaging of attenuation coefficients, and HU to LAC conversion. Finally, the formation of µmaps and PET difference measurements require accurate spatial co-registration of all data. All of these listed factors and experimental steps may lead to systematic errors in the obtained results, despite thorough planning and execution of the experimental study.

CONCLUSIONS
An optimized method for the conversion of CT to PET attenuation coefficients has been derived by systematic measurement of 38 different materials frequently used in PET/MR hybrid imaging. In contrast to established methods, the new conversion also considers highly attenuating materials such as electronic components and metals, thus improving attenuation correction of hardware components in PET/MR hybrid imaging.

CONFLICT OF INTEREST
The author Titus Lanz is an employee of Rapid Biomedical GmbH, Rimpar, Germany. All other authors declare that there are no conflicts of interest.

ETHICAL APPROVAL
This article does not contain any studies with human participants performed by any of the authors.