Center for Medical Physics and Biomedical Engineering, Medical University of ... Institute for MRI, University of DuisburgâEssen, Essen 45141, Germany; and ...
Quality control for quantitative multicenter whole-body PET/MR studies: A NEMA image quality phantom study with three current PET/MR systems Ronald Boellaarda) Department of Radiology and Nuclear Medicine, VU Medical Center, Amsterdam 1081 HV, The Netherlands; European Association of Nuclear Medicine Research Ltd., Vienna 1060, Austria; and European Association of Nuclear Medicine Physics Committee, Vienna 1060, Austria
Ivo Rausch and Thomas Beyer Center for Medical Physics and Biomedical Engineering, Medical University of Vienna, Vienna 1090, Austria
Gaspar Delso GE Healthcare and University Hospital of Zurich, Zurich 8091, Switzerland
Maqsood Yaqub Department of Radiology and Nuclear Medicine, VU Medical Center, Amsterdam 1081 HV, The Netherlands
Harald H. Quick Institute of Medical Physics, University of Erlangen-Nuremberg, Erlangen 91052, Germany; Erwin L. Hahn Institute for MRI, University of Duisburg–Essen, Essen 45141, Germany; and High Field and Hybrid MR-Imaging, University Hospital Essen, Essen 45147, Germany
Bernhard Sattler Department of Nuclear Medicine, University Hospital of Leipzig, Leipzig 04103, Germany and European Association of Nuclear Medicine Physics Committee, Vienna 1060, Austria
(Received 27 April 2015; revised 6 July 2015; accepted for publication 2 September 2015; published 22 September 2015) Purpose: Integrated positron emission tomography/magnetic resonance (PET/MR) systems derive the PET attenuation correction (AC) from dedicated MR sequences. While MR-AC performs reasonably well in clinical patient imaging, it may fail for phantom-based quality control (QC). The authors assess the applicability of different protocols for PET QC in multicenter PET/MR imaging. Methods: The National Electrical Manufacturers Association NU 2 2007 image quality phantom was imaged on three combined PET/MR systems: a Philips Ingenuity TF PET/MR, a Siemens Biograph mMR, and a GE SIGNA PET/MR (prototype) system. The phantom was filled according to the EANM FDG-PET/CT guideline 1.0 and scanned for 5 min over 1 bed. Two MR-AC imaging protocols were tested: standard clinical procedures and a dedicated protocol for phantom tests. Depending on the system, the dedicated phantom protocol employs a two-class (water and air) segmentation of the MR data or a CT-based template. Differences in attenuation- and SUV recovery coefficients (RC) are reported. PET/CT-based simulations were performed to simulate the various artifacts seen in the AC maps (µ-map) and their impact on the accuracy of phantom-based QC. Results: Clinical MR-AC protocols caused substantial errors and artifacts in the AC maps, resulting in underestimations of the reconstructed PET activity of up to 27%, depending on the PET/MR system. Using dedicated phantom MR-AC protocols, PET bias was reduced to −8%. Mean and max SUV RC met EARL multicenter PET performance specifications for most contrast objects, but only when using the dedicated phantom protocol. Simulations confirmed the bias in experimental data to be caused by incorrect AC maps resulting from the use of clinical MR-AC protocols. Conclusions: Phantom-based quality control of PET/MR systems in a multicenter, multivendor setting may be performed with sufficient accuracy, but only when dedicated phantom acquisition and processing protocols are used for attenuation correction. C 2015 American Association of Physicists in Medicine. [http://dx.doi.org/10.1118/1.4930962] Key words: PET/MR, attenuation correction, quantification, harmonization, multicenter, multivendor, quality control, NEMA image quality phantom
1. INTRODUCTION Multimodality imaging by means of positron emission tomography/magnetic resonance (PET/MR) imaging supports the quasi-simultaneous acquisition of molecular, functional, and structural information. Initially, several technical limitations 5961
Med. Phys. 42 (10), October 2015
had to be overcome in order to build PET/MR systems. First of all, conventional PET detectors use photomultiplier tubes (PMT), which cannot be used in a (strong) magnetic field.1–4 This technical challenge can be resolved by operating these PET components further away from the MR, such that the magnetic field strength is sufficiently low and cross talk effects
0094-2405/2015/42(10)/5961/9/$30.00
© 2015 Am. Assoc. Phys. Med.
5961
5962
Boellaard et al.: PET/MR quality control
are limited. In this way, state-of-the-art PET and MR systems can be combined into a single, coplanar system. This strategy was followed by Philips Healthcare and resulted in the Ingenuity TF PET/MR system.5,6 Another solution, as required for fully-integrated systems allowing for simultaneous PET and MR imaging, is to use PET detectors that are not affected by the magnetic field from the MR system.2,4 These detectors are based on either avalanche photodiodes (APD), as used by Siemens Healthcare in the Biograph mMR PET/MR system,7 or silicon photomultipliers (SiPM) as recently introduced by GE Healthcare in their GE SIGNA PET/MR system. All systems are presently commercially available and allow collecting PET and MR data in a single imaging session with diagnostic PET and MR imaging quality. For visual interpretation, images are of sufficient quality and combined PET/MR interpretation in selected clinical cases may provide more accurate diagnostic performance than that based on PET/CT.8–14 However, the quantitative accuracy of PET images in the context of PET/MR is less accurate than that of PET/CT. In PET/CT, the CT data can be used for the attenuation correction (AC) of PET because the CT images represent attenuation coefficients (Hounsfield units, HU), although at a different photon energy (80–140 keV) than the 511 keV annihilation photons. CT images can be rescaled to 511 keV linear attenuation coefficients (LAC) by using bilinear scaling methods.15,16 Yet, also use of CT may result in PET attenuation correction artifacts in case of patient motion or breathing, the presence of metal implants, and use of contrast agents. To avoid these artifacts, recommendations for PET/CT imaging have been published as to how these artifacts can be mitigated or reduced.17 In fact, the use of CT images for attenuation correction of the PET emission data in PET/CT studies has now become the reference for the validation of the quantitative performance of PET/MRI in the context of intraindividual comparison studies.8,18,19 In case of PET/MR imaging, the attenuation of the PET signal by patient tissues needs to be derived from the MR data rather from a CT image or a PET-based transmission scan. This is usually achieved by performing a dedicated and fast MR sequence followed by tissue classification and segmentation.20 Next, attenuation coefficients are assigned to the various tissue classes and segmentations. Although implemented for clinical use in all current available PET/MR systems, this procedure comes with a number of inherent challenges, which have been described extensively in the literature.21 In short, patient bed and MR radio frequency (RF) coils are not visible for the MR. Thus, CT-based templates of these objects need to be added to the attenuation coefficient image (µ-map). The µ-map may suffer from truncation when exceeding the spatial constraints of the MR field-of-view (FOV).22 In particular, for PET/MR imaging, the patient arms, which rest beside the patient body during PET/MR exams, are often not fully included in the MR FOV, thus, resulting in incorrect attenuation correction.22 In order to derive a µ-map for the chest, the lungs are typically segmented from the MR image and assigned a constant attenuation value. Finally, the implemented approaches for deriving the patient µ-map do not include bones or substitute bones with the LAC of soft tissue, thus, resulting in large negative Medical Physics, Vol. 42, No. 10, October 2015
5962
biases in or near bones and skull.23,24 Presently, several new methods are being developed and implemented to mitigate these various limitations of attenuation correction in PET/MR, such as the use of CT bone templates and atlas-based approaches (e.g., for the head/brain region), use of ultrashort echo time (UTE), or zero echo time (ZTE) MR sequences to visualize bones,25–28 neuronal-network based approaches aiming at continuous valued CT-like µ-maps,29,30 or new reconstruction methods that can estimate both emission and attenuation, the so called maximum likelihood reconstruction of activity and attenuation (MLAA) methods.31–33 In order to validate and calibrate the quantitative performance of PET (of the PET/CT and PET/MR), fluid fillable phantoms are typically used. These phantoms may contain one or more compartments that can be filled with solutions of different activity concentrations. A frequently used quality control (QC) phantom to assess the quantitative performance for PET/CT systems is the National Electrical Manufacturers Association (NEMA) NU 2 image quality phantom (NEMA IQ).34 This phantom (Fig. 1) is also used within the EARL FDG-PET/CT accreditation program to harmonize PET/CT system performance for quantitative multicenter studies.17 The composition of this phantom is different from that of human tissues, such that MR-AC and tissue segmentation algorithms may not be able to properly visualize these phantoms and to assign correct LAC for different fluids and the phantom housing. An example of potential errors and difficulties in generating a MR-based µ-map for a NEMA IQ phantom setup with PET/MR can be found in Ziegler et al.35 In order to achieve a homogeneous MR signal response from the water containing compartments, NaCl is added to the water solution, which changes the dielectric properties of the phantom fluid, thus, increasing signal homogeneity during MR imaging of these phantoms.35 Second, phantom compartment walls are made of plastic or synthetic materials, which are not visible in the MR image. As a consequence, phantom walls are systematically ignored in the MR-based µ-map, thus resulting in a negative bias in the reconstructed PET images. Moreover, the NEMA IQ phantom contains an 18 cm long,
F. 1. Picture of the NEMA IQ phantom with spheres and lung insert.
5963
Boellaard et al.: PET/MR quality control
5 cm diameter cylindrical plastic insert filled with polystyrene spheres mimicking the lungs. Both size and shape of this insert are far from that of lungs and may be completely missed during processing of the MR data into the µ-map. Thus, the use of phantoms designed and standardized for calibrating and characterizing the PET/CT systems may be challenging when used in the context of PET/MR imaging. It could not only hamper a proper calibration of a PET/MR system at a single site, but would potentially limit also the opportunity to harmonize the quantitative PET performance across different sites or systems. The purpose of the present study is, therefore, to test the applicability of different protocols for PET quality control in multicenter and multivendor PET/MR imaging. 2. MATERIAL AND METHODS 2.A. PET/MR systems and attenuation correction protocols
In the present study, three commercially available combined PET/MR systems were used. The systems included are a Philips Ingenuity TF PET/MR, the Siemens Biograph mMR, and the GE SIGNA PET/MR. At the time of measurements, the GE SIGNA PET/MR system was formally still a prototype. Yet clinical studies were performed to provide the necessary input data for FDA clearance and CE certification submission. Therefore, no changes with regard to the processing and reconstruction software were made to the current product version of the GE SIGNA PET/MR system that would affect the results described in this study. All systems were equipped with dedicated MR imaging protocols for acquiring and generating a patient µ-map (clinical protocol) as well as adapted MR imaging protocols for generating a µ-map for phantoms or specifically for the NEMA IQ phantom (phantom protocol).
5963
tion, which are segmented into three tissue classes (air, lungs, and soft tissue) that are assigned with fixed LAC for each tissue class.5 MR-AC acquisitions are performed using the built-in RF transmit/receive body coil of the MR system. The Siemens Biograph mMR system uses a Dixon-VIBE sequence in coronal orientation. In- and opposed-phase images are processed into fat- and water images, and the resulting images are segmented into four tissue classes (air, lungs, fat, and soft tissue), each assigned with a fixed LAC for 511 keV photons.36 The GE SIGNA PET/MR system performs a 3D, dualecho, RF spoiled gradient recalled echo (SPGR) sequence (LAVA-FLEX) in axial orientation. Next, water and fat images are derived from in-phase and out-of-phase (opposed-phase) images. The MR dataset is then classified into four tissue classes (air, lung, fat, and soft tissue). The air and lung tissue classes are binary, whereas the classification between fat and soft tissue—each with fixed attenuation values—is variable: the LAVA-Flex fat/water ratio is used to obtain a weighted average of fat (0.086 cm−1) and soft tissue (0.100 cm−1).37 The segmentation settings depend, however, on the anatomical region being identified during setup (localizer series) of the scanning procedure. As the NEMA IQ phantom consists mainly of a large compartment containing a water solution, the “hip” was selected as anatomical region. This selection results in the most straightforward MR-AC processing considering only air and soft tissue, thereby avoiding additional processing to deal with (small) air cavities and, e.g., chemotherapy ports, but at the cost of ignoring the presence of the lung insert, as will be discussed later. The MR data of the NEMA IQ phantom were acquired with the built-in RF transmit/receive body coil. All relevant protocol parameters for the clinical protocol are provided in Table I. 2.C. Phantom protocol
2.B. Clinical protocol
The clinical protocols used for MR-AC were as follows. The Philips Ingenuity TF PET/MR systems acquire transverse T1-weighted gradient echo (GRE) MR images in axial orienta-
The phantom protocols used for MR-AC in this study apply different approaches. The Philips Ingenuity TF PET/MR system performs a dedicated NEMA IQ MRI scan based on transverse T1w-GE but with longer scan duration, saturation bands, and a smaller voxel size (2×2×2 mm rather than 3×3×6 mm
T I. Imaging parameters of the clinical protocol used for MR-AC of the NEMA IQ phantom.
PET/MR system
Sequence
Number: classes
LAC (cm−1)
Phantom housing included?
Acquisition time MR (min)
Acquisition time PET (min)
Philips
2D GE axial
3: air, lungs, soft tissue
0.000 0.022 0.096
No
0:33
5:00
Siemens
3D Dixon VIBE coronal
4: air, lungs, soft tissue, fat
0.000 0.024 0.100 0.085
No
0:30–0:40
5:00
GEa
3D LAVA-FLEX
4: air, lungs, soft tissue, fat
0.000 0.018 0.100 0.086
No
0:18
5:00
a
Actual LAC is binary for lung and air and a weighted average of fat (0.086 cm−1) and water (0.100 cm−1) for the remaining voxels.
Medical Physics, Vol. 42, No. 10, October 2015
5964
Boellaard et al.: PET/MR quality control
5964
T II. Imaging parameters of the phantom protocol used in this study for MR-AC of the NEMA IQ phantom. NA = not applicable (CT template). PET/MR system
Sequence
Number: classes
LAC (cm−1)
Phantom housing included?
Acquisition time MR (min)
Acquisition time PET (min)
Philips Siemens GE
2D GE axial 3D Dixon VIBE coronal CT-based template of phantom
2: air, water 2: air, water NA
0.000, 0.096 0.000, 0.096 NA
Yes No Yes
18:47 0:30–0:40 NA
5:00 5:00 5:00
as used in the clinical protocol). Next, the resulting MR images are segmented into two classes only (water and air) as opposed to three classes (soft tissue, air, and lungs) of the clinical protocol. Fixed attenuations coefficients are then assigned to these classes (water = 0.096 cm−1, air = 0.0 cm−1). The phantom housing is considered by extending the outer contour seen on the MR data by 2 mm.5 The Siemens Biograph mMR system applies a similar two-class approach implemented for general PET phantom imaging (at the time of collecting the data, i.e., in September 2013, software version VB18P with no dedicated NEMA IQ procedure available). After collecting the MR data by using the same imaging protocol for MR-based attenuation correction as applied with the clinical protocol, the data are segmented into only two classes (water and air) and fixed attenuation coefficients are assigned to these phantom classes (water = 0.096 cm−1, air = 0.0 cm−1). The phantom housing is not considered in this approach. The GE SIGNA PET/MR system applies a CT-based template for this phantom. Attenuation coefficients for the phantom housing and its fluid fillings are derived from a predefined µ-map based on CT data. The µmap is registered to the nonattenuation corrected reconstructed time-of-flight (TOF)-PET image and subsequently used during PET reconstruction with attenuation correction. In all cases and on all three systems, attenuation from the patient table is taken into account by the use of predefined templates. All relevant protocol parameters for the phantom protocol are provided in Table II. 2.D. Experiments
In this study, a NEMA IQ phantom (Fig. 1) was imaged on all three different PET/MR systems following the specifications of the EARL PET/CT program.17 The NEMA IQ phantom was filled such as to obtain a FDG activity concentration of 2 kBq/ml in the large background compartment (volume 9.4 l) and 20 kBq/ml in each of the six spheres resulting in a sphere-to-background ratio of 10:1. To fill the compartments of the phantom, rather than using pure water, a solution of 50%
water and 50% physiological saline solution (0.9%) was used before adding the FDG to these solutions. Next, the phantom was scanned using a PET acquisition duration of 5 min in combination with both the clinical and phantom protocols for MR-AC. Finally, all PET data were reconstructed using the clinical and phantom protocols derived µ-maps and using image reconstruction settings as approved by EARL for their equivalent PET/CT counterparts. These reconstruction settings are listed in Table III. All other corrections needed to obtain quantitative PET images, such as corrections for randoms, scatter, dead time, decay, and variation in detector efficiencies, were applied as is standard in the three systems. 2.E. Data analysis
The µ-maps generated using the clinical and phantomspecific protocols applied on the three different systems were inspected for artifacts and for assignment of incorrect attenuation coefficient values. In addition, the reconstructed PET data were analyzed using a standardized analysis tool as used and provided by EARL.38 This tool semiautomatically places a predefined volume of interest (VOI) template onto the image data set. The template contains six regions of interest (ROIs) with varying sizes corresponding to that of the spheres. In addition, ten regions of interest each with 2.5 cm diameter are placed in the background compartment such that these do not overlap with those of the spheres. After alignment of the template ROI with the spheres and phantom, the tool determines the value and location of the voxel with the maximum activity concentration per sphere. Next, for each sphere, a 3D region growing is started using 50% of the maximum activity concentration value per sphere, corrected for local background, to generate a 3D VOI. Finally, for each sphere, the average and maximum activity concentration within these VOIs are derived and compared with the known activity concentration based on the phantom filling parameters. The ratio of observed activity concentration and expected activity concentration is referred
T III. PET image reconstruction settings applied for the phantom scans on the various PET/MR systems. TOF = time of flight, OSEM = ordered subset expectation maximization, PSF = point spread function (resolution recovery), OP = ordinary Poisson, NA = not available (i.e., cannot be seen nor defined by user).
Method
Iterations, subsets
Matrix size
TOF-OSEM (“blob-os-tf”) OP-OSEM TOF–PSF-OSEM
NA 3i,21s 4i,28s
288 × 288 256 × 256 256 × 256
PET/MR system Philips Ingenuity TF Siemens Biograph mMR GE SIGNA (prototype)
Medical Physics, Vol. 42, No. 10, October 2015
(Post) reconstruction filter None Gaussian, 5 mm FWHM Gaussian, 7 mm FWHM
5965
Boellaard et al.: PET/MR quality control
to as the recovery coefficient (RC) throughout the remainder of the paper. RCs will be shown as function of sphere size and compared with the EARL RC specifications. The average activity concentration seen within the background ROIs is compared with the expected background compartment activity concentration and used to verify the calibration accuracy of the systems. 2.F. Simulations
Noise free simulations were performed to provide a standard of reference and to verify and understand the quantitative biases seen in the PET images of the NEMA IQ phantom, as will be shown later. Input data were derived from performing the NEMA IQ phantom study in exactly the same manner using a PET/CT system (Philips Gemini TF). The obtained PET emission image and µ-map derived from the low-dose CT image were used to simulate TOF and non-TOF emission projection data including the effects of attenuation. TOF was simulated for both 325 and 650 ps timing resolution, thereby closely simulating the TOF performances of the GE and Philips PET/MR systems, respectively. The setup and procedure of the simulations are identical to those described in detail by Boellaard et al.33 In the context of this study, simulations were designed to study the artifacts seen in the µ-maps generated using the clinical and phantom-specific protocols. After generating the emission projection data, both TOF and non-TOF reconstructions based on ordered subset expectation maximization (OSEM) were performed since it was shown that PET image artifacts resulting from MR-AC errors are affected by TOF.33,39,40 TOF and non-TOF reconstructions were per-
5965
formed using an image matrix size of 256 × 256, an isotropic voxel size of 2.5 mm, 6 iterations, and 16 subsets. During the PET emission reconstruction, several different µ-maps were used to correct for attenuation: (1) the correct µ-map including attenuation from the phantom housing and the lung insert; (2) a µ-map where the outer phantom wall is missing and the lung insert attenuation coefficient is set to 0.0 cm−1 (air); and (3) a µ-map where the outer phantom wall is missing and the lung insert attenuation coefficients are set to 0.096 cm−1 (water).
3. RESULTS 3.A. Experimental results
An overview of obtained µ-maps of the NEMA IQ phantom using the three PET/MR systems and the clinical or phantomspecific µ-maps protocols is given in Fig. 2. Observations when using clinical MR-AC protocols are that for the Philips Ingenuity TF and GE SIGNA PET/MR systems, the lung insert was missed out and set to equal to the attenuation coefficient of the large background compartment. Of note, hip was used to indicate the “anatomical” coverage of the MR-AC scan with the GE SIGNA PET/MR system. Moreover, both the Siemens Biograph mMR and GE SIGNA PET/MR systems considered the large water/saline filled background compartment in the NEMA IQ phantom as fat and consequently assigned an incorrect attenuation coefficient (0.085 and 0.086 cm−1, respectively, as opposed to the expected 0.096 cm−1 for water-based solutions). When using the phantom-specific protocols, the lung insert was seen for all systems, but both the Philips Ingenuity TF
F. 2. Overview of µ-maps obtained using the various systems (columns) using the clinical [(a)–(c)] or phantom [(d)–(f)] protocols. µ-values for the lung insert are provided and indicated by the red arrows. Please note that all systems apply a template to take attenuation from the bed into account during PET image reconstruction, but not all systems show the bed template in the generated µ-maps [(a) and (d)]. On the Philips system, a phantom holder made from nonattenuating foam was used during the phantom MR-AC protocol explaining the small elevation of the phantom (e). Medical Physics, Vol. 42, No. 10, October 2015
5966
Boellaard et al.: PET/MR quality control
5966
T IV. Linear attenuation coefficients (cm−1) observed in the IQ phantom using the various protocols. Background compartment
Lung insert
Philips Siemens GE
Clinical protocol
Phantom protocol
Clinical protocol
Phantom protocol
0.096 0.022 0.086
0.000 0.000 0.032
0.096 0.085 0.086
0.096 0.096 0.096
3.B. Simulations
Simulations were performed to serve as standard of reference to verify and explore the quantitative biases that may be seen in case of imperfect or incorrect µ-maps. Moreover, the impact of incorrect µ-maps on the reconstructed PET images may be affected by TOF capabilities of the system. Figure 5 illustrates the biases that can be expected in a phantom setup when using clinical protocols.
and Siemens Biograph mMR PET/MR systems assigned an attenuation coefficient of 0.0 cm−1 (air) (as opposed to the correct 0.032 cm−1 for the lung insert). In all cases, the water equivalent attenuation coefficient was assigned to the large water-filled compartment. An overview of observed linear attenuation coefficients for the clinical and phantom protocols is given in Table IV. Figure 3 illustrates the ratio of average measured and expected activity concentration in the background compartment of the NEMA IQ phantom. Large underestimations were observed when using the clinical protocol on the GE SIGNA PET/MR and Siemens Biograph mMR. Inspection of µ-maps revealed that attenuation coefficients corresponded to that of fat (Table IV). Using the phantom-specific protocols instead, the results improved to within expected levels (
