Home

Search

Collections

Journals

About

Contact us

My IOPscience

Evaluation and automatic correction of metal-implant-induced artifacts in MR-based attenuation correction in whole-body PET/MR imaging

This content has been downloaded from IOPscience. Please scroll down to see the full text. 2014 Phys. Med. Biol. 59 2713 (http://iopscience.iop.org/0031-9155/59/11/2713) View the table of contents for this issue, or go to the journal homepage for more

Download details: IP Address: 128.193.164.203 This content was downloaded on 28/05/2014 at 09:41

Please note that terms and conditions apply.

Institute of Physics and Engineering in Medicine Phys. Med. Biol. 59 (2014) 2713–2726

Physics in Medicine and Biology

doi:10.1088/0031-9155/59/11/2713

Evaluation and automatic correction of metal-implant-induced artifacts in MR-based attenuation correction in whole-body PET/MR imaging G Schramm 1,2 , J Maus 1 , F Hofheinz 1 , J Petr 1 , A Lougovski 1 , B Beuthien-Baumann 3 , I Platzek 4 and J van den Hoff 1,3 1 Helmholtz-Zentrum Dresden-Rossendorf, Institute of Radiopharmaceutical Cancer Research, Department of Positron-Emission-Tomography, Dresden, Germany 2 Technische Universit¨at Dresden, Faculty of Medicine Carl Gustav Carus, Dresden, Germany 3 University Hospital Carl Gustav Carus, Department of Nuclear Medicine, Dresden, Germany 4 University Hospital Carl Gustav Carus, Department of Radiology, Dresden, Germany

E-mail: [email protected] Received 11 December 2013, revised 18 March 2014 Accepted for publication 25 March 2014 Published 6 May 2014 Abstract

The aim of this paper is to describe a new automatic method for compensation of metal-implant-induced segmentation errors in MR-based attenuation maps (MRMaps) and to evaluate the quantitative influence of those artifacts on the reconstructed PET activity concentration. The developed method uses a PETbased delineation of the patient contour to compensate metal-implant-caused signal voids in the MR scan that is segmented for PET attenuation correction. PET emission data of 13 patients with metal implants examined in a Philips Ingenuity PET/MR were reconstructed with the vendor-provided method for attenuation correction (MRMaporig , PETorig ) and additionally with a method for attenuation correction (MRMapcor , PETcor ) developed by our group. MRMaps produced by both methods were visually inspected for segmentation errors. The segmentation errors in MRMaporig were classified into four classes (L1 and L2 artifacts inside the lung and B1 and B2 artifacts inside the remaining body depending on the assigned attenuation coefficients). The average relative SUV av ) between PETorig and PETcor of all regions showing wrong differences (εrel attenuation coefficients in MRMaporig were calculated. Additionally, relative SUVmean differences (εrel ) of tracer accumulations in hot focal structures inside or in the vicinity of these regions were evaluated. MRMaporig showed erroneous attenuation coefficients inside the regions affected by metal artifacts and inside 0031-9155/14/112713+14$33.00

© 2014 Institute of Physics and Engineering in Medicine Printed in the UK & the USA 2713

G Schramm et al

Phys. Med. Biol. 59 (2014) 2713

the patients’ lung in all 13 cases. In MRMapcor , all regions with metal artifacts, except for the sternum, were filled with the soft-tissue attenuation coefficient and the lung was correctly segmented in all patients. MRMapcor only showed av (mean ± standard small residual segmentation errors in eight patients. εrel deviation) were: (−56 ± 3)% for B1, (−43 ± 4)% for B2, (21 ± 18)% for L1, (120 ± 47)% for L2 regions. εrel (mean ± standard deviation) of hot focal structures were: (−52 ± 12)% in B1, (−45 ± 13)% in B2, (19 ± 19)% in L1, (51 ± 31)% in L2 regions. Consequently, metal-implant-induced artifacts severely disturb MR-based attenuation correction and SUV quantification in PET/MR. The developed algorithm is able to compensate for these artifacts and improves SUV quantification accuracy distinctly. Keywords: PET/MR, attenuation correction, metal artifacts S Online supplementary data available from stacks.iop.org/PMB/59/2713/mmedia

(Some figures may appear in colour only in the online journal)

1. Introduction Attenuation correction still remains one of the key challenges in hybrid whole-body PET/MR imaging (Hofmann et al 2009) affecting the quantitative accuracy of the reconstructed PET activity distribution (Keereman et al 2011, Schramm et al 2013a). Since in both currently available whole-body PET/MR systems (Zaidi et al 2011, Delso et al 2011) attenuation correction is based on MR image segmentation and tissue type classification (Martinez-M¨oller et al 2009, Hu et al 2009, Schulz et al 2011), MR signal artifacts distort the attenuation correction algorithms. Especially in patients with metal implants such as endoprotheses, sternal cerclages or dental fillings (Keller et al 2013, Ladefoged et al 2013), signal voids in the MR scans caused by susceptibility artifacts (Schenk 1996) (see figure 1(a)) result in erroneously segmented attenuation maps. Metal implants are present in a substantial fraction of performed PET/MR investigations. For instance, at our institution, 19 out of 316 PET/MR whole-body investigations performed during the last six months were affected by metal implants leading to artifacts in the MR-based attenuation maps (MRMaps). In contrast to PET/CT, where metal implants cause streak artifacts (see figure 1(b)) that overestimate the attenuation as well as the reconstructed PET activity concentration (Goerres et al 2002, 2003, Bockisch et al 2004, Sureshbabu and Mawlawi 2005), metal-implant-induced signal voids in MR scans lead to an underestimation of the attenuation and activity concentration in PET/MR (Keller et al 2013, Ladefoged et al 2013). Ladefoged et al (2013) analyzed PET/MR examinations of four patients with endoprotheses and found severe underestimations of the reconstructed activity concentrations in regions affected by metal-implant-induced signal voids in the MR scan. To compensate for these artifacts, Ladefoged et al (2013) proposed a semi-automatic algorithm which fills the voids caused by endoprotheses in the attenuation map with the soft-tissue attenuation coefficient (AC) by inpainting. Using this algorithm, it was shown that the bias in the activity concentration could be partly corrected. However, since this method requires time consuming user interaction in every case, its applicability in clinical routine is questionable. In order to overcome this limitation, we have developed an improved algorithm which compensates the metal-implant-induced artifacts in MRMaps without requiring any user interaction. 2714

G Schramm et al

Phys. Med. Biol. 59 (2014) 2713

(a)

(b)

Figure 1. MRI signal voids (a) and CT streak artifacts (b) caused by endoprotheses of

patient 10. Note that the extent of the signal voids in the MRI caused by dephasing is much bigger than the actual extent of the endoprotheses. Table 1. Summary of patients included in the analysis. The columns show the patient number, the used radiotracer and the location of the metal implants.

Patient

Tracer

Location of metal implant

1 2 3 4 5 6 7 8 9 10 11 12 13

[18 F]-FDG [18 F]-NaF [18 F]-NaF [18 F]-FDG [18 F]-NaF [18 F]-FDG [18 F]-FDG [18 F]-FDG [18 F]-FDG [18 F]-NaF [18 F]-FDG [18 F]-NaF [18 F]-NaF

sternum left and right hip, right knee left and right knee left and right hip sternum sternum sternum, left knee sternum sternum left and right hip sternum left hip, left knee hip, symphysis

In this article we pursue two objectives: first, we describe a new algorithm to compensate for metal-implant-induced segmentation artifacts and evaluate the resulting attenuation maps by visual inspection. Second, we investigate the quantitative influence of metalimplant-induced attenuation errors on the reconstructed activity concentration in PET/MR examinations. 2. Materials and methods 2.1. Data acquisition and image processing

To evaluate the influence of metal-implant-induced segmentation errors in MRMaps we have reconstructed PET emission data of 13 oncological patients with two different methods for attenuation correction that are explained below. All patients were examined in an Ingenuity PET/MR system (Philips, Best, Netherlands) (Zaidi et al 2011). Metal implants causing susceptibility artifacts leading to signal voids in the MR scans were present in all patients. Patient details and a description of their metal implants are given in table 1. A standard 3D T1-weighted gradient echo sequence, called atMR scan, (flip angle 10◦ , TE 2.3 ms, TR 4 ms, voxel size 3 mm × 3 mm × 6 mm) was acquired for each patient and used in the vendor-provided and in the developed segmentation algorithm, respectively. In all cases, the PET emission data were acquired for 2 min per bed position. All examinations were performed in free-breathing mode. 2715

G Schramm et al

Phys. Med. Biol. 59 (2014) 2713

contour completion

atMR scan

MRMaporig

MRMapcor

PETorig

PETcor

PET emission data

Figure 2. Workflow of PET image reconstruction. (1) PETorig is reconstructed using the vendor-generated MRMaporig which contains segmentation errors due to metal artifacts. (2) MRMapcor is generated using segmentation of the atMR scan and contour completion with PETorig . (3) PETcor is reconstructed using the same PET emission data as in (1), but MRMapcor for attenuation correction.

As shown in figure 2, we first generated the standard MRMaporig using the vendorprovided implementation of a three class segmentation (Hu et al 2009) (air μ = 0 cm−1 , lung tissue μ = 0.022 cm−1 and soft tissue μ = 0.096 cm−1 ) to reconstruct PETorig . In all cases, MRMaporig contained segmentation errors. Surprisingly, the segmentation errors did not only occur at the locations of the metal implants, but also in the remote lung region. Second, we applied the algorithm for metal artifact reduction developed by our group to generate MRMapcor yielding PETcor . A detailed description of the algorithm is given in subsection 2.3. For PET image reconstruction, a listmode ordered-subset expectationmaximization algorithm with 33 subsets and three iterations was used in all cases. No post filtering in the image space was applied. The exact same setting is used in clinical whole-body reconstructions. For patient 10, where a CT scan from a previous examination was available, we were able to create a third attenuation map (MRMapmetal yielding PETmetal ) which contains the actual extent of the endoprotheses. The delineation of the endoprotheses was done in the coregistered CT scan using the software ROVER (ABX, Radeberg, Germany). In MRMapmetal we used an attenuation coefficient of 0.36 cm−1 for the endoprotheses made of a titanium alloy (Berger et al 2010). We decided to use this fixed literature value for titanium because, as shown in Burger et al (2002), the bilinear transformation from Hounsfield units to 511 keV attenuation coefficients can be erroneous for high Hounsfield units. 2.2. Classification of metal artifacts

After visual inspection of the atMR scan and MRMaporig of all patients we divided all segmentation errors in MRMaporig caused by metal implants into four classes (L1, L2, B1 and B2). The classes L1 and L2 contain segmentation errors in the lung region and classes B1 and B2 contain artifacts in the remaining body. A description of the classes is given in table 2 and examples of all artifacts are shown in figure 3. 2.3. Description of the developed algorithm for metal artifact reduction

Based on a delineation of PETorig and the atMR scan, our algorithm consists of the following steps (see figure 4). 2716

G Schramm et al

Phys. Med. Biol. 59 (2014) 2713

Table 2. Classification of segmentation errors in the MRMap caused by metal artifacts.

Artifact Description

Origin

Patient

Example

L1

signal void from metal implant in sternum connecting lung to background region signal void from bilateral metal implants segmented as lung signal voids from unilateral metal implants signal voids from bilateral endoprotheses segmented as lung

1, 5, 6, 7, 8, 9, 11, 13 2, 3, 4, 10, 12 2, 3, 7, 12, 13 2, 4, 10

Figure 3(b)

L2 B1 B2

air and soft-tissue attenuation coefficient (0 cm−1 , 0.096 cm−1 ) in lung tissue soft-tissue attenuation coefficient (0.096 cm−1 ) in lung tissue air attenuation coefficient (0 cm−1 ) in soft tissue lung tissue attenuation coefficient (0.022 cm−1 ) in soft tissue

Figure 3(a) Figure 3(c) Figure 3(a)

(i) The body contour of the patient (CPET ) is delineated in PETorig using an automatically determined background intensity threshold as described in Schramm et al (2013b) (see figure 4(b)). (ii) A global background intensity threshold (IBG,atMR ) in the atMR scan corresponding to the first local minimum of the atMR voxel intensity histogram is calculated (see figure 5(a)). (iii) All voxels in the atMR scan belonging to CPET and having an MR intensity below IBG,atMR are set to the maximum MR intensity. In this way, cavities in the patient contour of the atMR originating from metal-induced signal voids are closed (see figure 4(c)). (iv) The closed MR contour is filled with the soft-tissue attenuation coefficient (0.096 cm−1 ) resulting in a binary attenuation map (background and soft tissue). (v) The axial position of the patient’s lung is calculated using an automatic detection of the transaxial slice containing the patient’s shoulder. This slice is determined by analyzing the average MR intensity per transaxial slice in the atMR scan (see figure 5(b)). The volume located between the transaxial shoulder slice and the transaxial slice 35 cm proximal to the shoulder slice is considered as subvolume potentially containing the lung (Vlung ). If the shoulder is not included in the scan, the topmost slice is used as the upper boundary for the subvolume. (vi) Inside Vlung , all lung voxels are delineated using an automatically determined upper intensity threshold (Ilung ) which is derived from the MR intensity histogram of all voxels inside Vlung (see figure 5(c)). (vii) The MRMap is smooth·ened using a Gaussian filter with FWHM 4 mm. (viii) The attenuation template of the patient bed is added to MRMapcor (see figure 4(d)). The algorithm was implemented in C++ and has a computation time of approximately 1 min on a single core AMD Opteron 6376 CPU with 1.4 GHz for a whole-body PET/MR investigation. 2.4. Image analysis

To evaluate the quantitative influence of metal-implant-induced segmentation errors in PETorig we analyzed two different quantities. First, we manually delineated three-dimensional regions with erroneous attenuation coefficients in MRMaporig . Inside these regions, we calculated the av ) using a voxel-wise linear model regression between average relative SUV difference (εrel PETorig and PETcor . An example for such a regression analysis is shown in figure 6. Second, tracer accumulations in hot focal structures within or in the close vicinity of regions with erroneous attenuation coefficients were delineated. These structures include tumors, degenerative uptake (e.g. fractures in NaF investigations) and non-pathological uptake 2717

G Schramm et al

Phys. Med. Biol. 59 (2014) 2713

(a) Patient 2

(b) Patient 7

L2

L1

B2

B2

B1 B1 (c) Patient 3

(d) Patient 1

L2

L1 B1

B1

Figure 3. Four examples of patients with metal implants and incorrect MRMaps. In each panel: (left) atMR scan with metal artifacts, (middle) MRMaporig and (right) MRMapcor . (a) Patient 2 with metal implants in the hip and right knee and incorrect attenuation coefficients in the lung, hip and knee. (b) Patient 7 with metal implants in the left knee and incorrect attenuation coefficients in the lung and left knee. (c) Patient 3 with metal implants in both knees and incorrect attenuation coefficients in the lung and both knees. (d) Patient 1 with metal in sternum and incorrect attenuation coefficients in the lung. The type of artifact as classified in table 2 is displayed in the panel showing MRMaporig .

(e.g. myocardium and bladder). The volume of interest (VOI) delineation was performed in ROVER (Hofheinz et al 2012) which uses adaptive thresholding and subtraction of local background. All VOIs were delineated in PETcor and subsequently applied to PETorig . For all VOIs absolute: mean εabs = SUVmean orig − SUVcor

(1)

and relative: εrel =

SUVmean

mean SUVmean orig − SUVcor

(2)

SUVmean cor differences were calculated. 2718

G Schramm et al

Phys. Med. Biol. 59 (2014) 2713

Patient 1 (a)

(b)

(c)

(d)

linear attenuation coefficient

Patient 2 (f)

(e)

(g)

(h)

(i)

air 0 cm-1 lung tissue 0.022 cm-1 soft tissue 0.096 cm-1

(j)

Figure 4. Corresponding transaxial (upper) and coronal slices (lower panel) of patients 1 and 2, respectively. (a), (f) PETorig , (b), (g) delineated patient contour CPET in PETorig . (c), (h) atMR scan with metal artifacts. The patient’s contour in the atMR was closed by CPET shown as the white line. (d), (i) MRMapcor . (e), (j) MRMaporig showing incorrect tissue class segmentation due to metal artifacts in the atMR scan. Moreover, truncation artifacts in the arms are present.

3. Results 3.1. Evaluation of artifacts in MRMaporig and MRMapcor

Table S1 (available at stacks.iop.org/PMB/59/2713/mmedia) shows a detailed overview of the observed artifacts in MRMaporig and MRMapcor caused by metal-implant-induced signal voids in the atMR scan. Examples for all types of artifacts that occurred in MRMaporig are shown in figure 3. In all 13 patients, the lung is incorrectly segmented in MRMaporig . In eight cases artifact L1 (lung partly filled with air and soft-tissue attenuation coefficient) and in five cases artifact L2 (lung completely filled with soft-tissue attenuation coefficient) is present. In all L1 cases except for patient 13, a signal void in the sternum is present connecting the lung region with low atMR signal to the outside (see figure 3(d)). 2719

G Schramm et al

Phys. Med. Biol. 59 (2014) 2713

(a)

(b)

(c)

Figure 5. (a) Logarithmic frequency distribution of MR signal intensity of the complete

atMR scan. The dashed line shows the chosen global background intensity threshold IBG,atMR . (b) Number of voxels N with an MR signal intensity greater than 75% of the mean signal intensity per transaxial slice. The dashed lines show the borders of the designated lung slices. The distal end of the designated lung slices is the slice where N reaches 50% of it’s maximum. (c) Logarithmic frequency distribution of MR signal intensity of designated lung slices in atMR scan. The dashed line shows the upper lung intensity threshold Ilung . Please note that the scale of the x-axis in (a) and (c) is different because of the arbitrary units.

(a)

(b)

5

(c)

(d) 8 SUVorig

SUV

6

300 R^2=0.97 m=0.61 200

4 100

2

0

0 0

2

4 6 SUVcor

8

0

Figure 6. Representative coronal slices of patient 2: (a) PETorig , (b) PETcor , (c) MRMaporig and (d) two-dimensional histogram of activity concentrations of all voxels in the VOI enclosed by the blue contour in (a)–(c). This VOI contains all voxels having a wrong attenuation coefficient caused by the left endoprothesis. In the two-dimensional histogram, the line of identity and the regression slope are shown as dashed and solid lines, respectively. The average relative SUV difference is calculated as m − 1, where m is the regression slope.

Moreover, five patients show artifact B1 (air attenuation coefficient inside the body) and three patients show artifact B2 (lung tissue attenuation coefficient inside the body). In the three patients with artifact B2, bilateral atMR signal voids caused by hip implants are wrongly segmented as lung (see figure 3(a)). The correct lung region was filled with the soft-tissue attenuation coefficient in these cases. In contrast to MRMaporig , the lungs of all patients are segmented correctly in MRMapcor . In six patients with metal implants in the sternum, the sternum is wrongly filled with the lung attenuation coefficient (see figure 3(d)). All regions inside the body showing artifacts B1 or B2 are filled with the soft-tissue attenuation coefficient. In patients 2 and 3, part of the surface of the knees (about 50 ml and 400 ml, respectively) is not completely recovered (see figure 4 (d)). In patient 2, a small part of the spine with low MR signal was filled with the lung tissue attenuation coefficient. In patient 8, abdominal air cavities were incorrectly assigned with the lung tissue attenuation coefficient. 2720

G Schramm et al

Phys. Med. Biol. 59 (2014) 2713

0.36

(b)

(c)

(d)

(e)

(f)

5

+70% (g)

(h)

(i)

0 2600

µ (1/cm)

(a)

HU

SUV

0

-70%

-466

Figure 7. Representative coronal slices of (a) MRMaporig , (b) MRMapcor , (c) MRMapmetal , (d) PETorig , (e) PETcor , (f) PETmetal , (g) relative difference between (d) and (f), (h) relative difference between (e) and (f) and (i) coregistered CT scan from a previous examination of patient 10. The endoprotheses in (c) were delineated in (i) and assigned with a fixed attenuation coefficient (0.36 cm−1 ). The relative difference in (g) and (h) are only shown for voxels with SUVmetal > 0.5 to achieve better visibility.

3.2. Influence of artifacts on the reconstructed activity distribution

Figure 7 shows PET images (middle row) of patient 10 reconstructed with MRMaporig , MRMapcor and MRMapmetal (top row). Additionally, relative difference images of PETorig and PETcor with PETmetal as a reference are shown in the bottom row. Compared to PETmetal , PETorig shows strongly decreased tracer uptake in a 2–3 cm thick halo around both endoprotheses. In contrast, PETcor only underestimates the activity distribution in a small region (about 0.5 cm thick) next to the surface of the endoprotheses. Additionally, PETorig and PETcor overestimate the activity distribution of voxels with low uptake between the endoprotheses. The average SUV differences in the region where MRMaporig shows erroneous attenuation coefficients are: −48% between PETorig and PETmetal and −13% between PETcor and PETmetal . Figure 8 shows MRMaporig , MRMapcor , PETorig , PETcor and the absolute (εabs ) as well as the relative SUV difference (εrel ) for patient 2. Artifact L2 (lung filled with soft-tissue attenuation coefficient in MRMaporig ) leads to an SUV overestimation of all voxels in the lung and its vicinity in PETorig . In contrast, artifact B1 (knee filled with air attenuation coefficient in MRMaporig ) and artifact B2 (regions around endoprotheses filled with lung attenuation coefficient in MRMaporig ) lead to an SUV underestimation in PETorig . av Table 3 summarizes the average relative SUV difference (εrel ) of all regions containing erroneous attenuation coefficients caused by metal artifacts in the atMR scan. Table S2 (available at stacks.iop.org/PMB/59/2713/mmedia) shows that compared to PETcor , PETorig underestimates the SUVmean of all tracer accumulations in hot focal structures in B1 and B2 regions, whereas it overestimates the SUVmean of all hot focal structures in L1 and L2 regions. The observed deviations are (mean ± standard deviation): −52% ± 12% in B1 regions, −45% ± 13% in B2 regions, 19% ± 19% in L1 regions, 51% ± 31% in L2 regions. 2721

G Schramm et al

Phys. Med. Biol. 59 (2014) 2713

(d)

(e)

5

(f)

(g)

5

+100%

rel

0.12

abs

(c)

SUV

(b)

µ (1/cm)

(a)

0

0

-5

-100%

Figure 8. Representative coronal slices of patient 2. (a) atMR scan (b) MRMaporig (c) MRMapcor (d) PETorig (e) PETcor (f) absolute and (g) relative difference between (d) and (e). In (g) the relative difference is only shown for voxels with SUVcor > 0.05 to improve rendering of the relevant structures. av Table 3. Average relative SUV differences (εrel ) between PETorig and PETcor in regions with erroneous attenuation coefficients caused by metal artifacts in the atMR scan. The mean and standard deviation for each class of artifact (see table 2) is given separately.

Patient

Region

Artifact

av εrel (%)

Patient

Region

Artifact

2 3 3 7 12 13 13

right knee left knee right knee left knee left knee bladder, symphysis hip

B1 B1 B1 B1 B1 B1 B1 mean sd L1 L1 L1 L1 L1 L1 L1 mean sd

−59 −58 −58 −55 −56 −50 −54 −56 3 32 48 15 1 −2 15 33 21 18

2 2 4 4 10 102

left endoprothesis right endoprothesis left endoprothesis right endoprothesis left endoprothesis right endoprothesis

B2 B2 B2 B2 B2 B2

−39 −37 −49 −46 −46 −43

lung lung lung lung lung

mean sd L2 L2 L2 L2 L2

−43 4 91 95 125 199 88

mean sd

120 47

1 5 6 8 9 11 13

lung lung lung lung lung lung lung

2 3 4 10 12

av εrel (%)

4. Discussion Signal voids due to susceptibility artifacts caused by metal implants lead to severe segmentation errors in MRMaps. On one hand, they can directly distort the segmentation when these signal voids are erroneously classified as air cavities in the MRMap. This type of artifact is scanner independent and affects MRMaps on all PET/MR systems. On the other hand, metal-implantcaused signal voids indirectly disturb the model-based Philips-provided lung segmentation as observed in all patients with L1 (air and soft-tissue AC in the lungs) or B2 (lung AC in the body) artifacts (see table 2 for artifact classification). Signal voids of the atMR scan in the sternum connect the lung region with low atMR signal to the outside. Thus, the search 2722

G Schramm et al

Phys. Med. Biol. 59 (2014) 2713

for connected regions with low signal inside the body that are supposed to be the lungs fails. Artifacts L1 and B2 strongly depend on the implementation of the segmentation algorithm and thus might be different on other systems. The metal-implant-induced shortcomings of MRMaporig are overcome by the developed algorithm which closes the contour in the atMR with a contour delineated in PETorig . As suggested in Ladefoged et al (2013), the algorithm fills the signal voids with the lung or softtissue attenuation coefficient depending on their location. In contrast to the semi-automatic algorithm described by Ladefoged et al (2013), our algorithm works fully automatic without requiring any user interaction. On top of that, the algorithm is not only capable of automatically filling cavities in the MRMap surrounded by soft tissue. It is also able to complete the patient’s outer contour and thus to fill cavities that are connected to the background. This is especially important for surface-near metal implants such as sternal cerclages or knee endoprotheses. As discussed in Ladefoged et al (2013), filling of the signal void regions caused by metal implants with a metal attenuation coefficient is not reasonable because the actual extent of the signal void region is much bigger than the extent of the metal implant. Ladefoged et al (2013) could show that a simple complete filling of the signal void regions with a metal attenuation coefficient leads to a severe SUV overestimation (>460% in SUVmax ). Moreover, generation of an MRMap including the correct attenuation information of the metal implant (e.g. via using a template for the implant) is difficult because the exact position of the implant in the signal void is not detectable in the atMR scan. As described in Goerres et al (2003), a slightly misaligned endoprothesis in the attenuation map can lead to artificial hot structures in the reconstructed PET image. Nonetheless, it is clear that a simple complete filling with the soft-tissue attenuation coefficient underestimates the attenuation of the metal implant and thus leads to an underestimation of the reconstructed SUVs. However, as suggested in the conclusion of Ladefoged et al (2013), the bias introduced by this method is small in FDG examinations and thus this method is suitable as a first-line correction method. With the different reconstructions of patient 10 (see figure 7), we could estimate that the SUV underestimation introduced by the soft-tissue filling is about −13% on average. Interestingly, a positive bias in regions with low tracer uptake close to the endoprotheses is observed which is probably due to the influence of the attenuation map on the scatter correction. In the absence of a ground truth for the correct attenuation map, we have calculated the SUV differences between PETorig and PETcor in order to estimate the quantitative influence of segmentation errors caused by metal artifacts on the reconstructed SUVs. As presented in tables 3 and S2 (available at stacks.iop.org/PMB/59/2713/mmedia), the average SUV differences in the regions with wrong attenuation coefficients agree well with the SUVmean differences found in hot focal structures inside or in the vicinity of these regions. Ladefoged et al (2013), report an SUVmax difference (mean ± standard deviation) of (−31±16)% between PETorig and PETcor in endoprotheses of four patients. This value is slightly but not significantly lower (p = 0.13 two-sample Kolmogorov–Smirnov test) than the value of (−43 ± 4)% found in three patients (B2 artifact) in this study. In B1 regions (air AC in soft-tissue), we found an average SUV underestimation in PETorig of (−56 ± 3)%. Bezrukov et al (2013) report an SUV underestimation of (−54 ± 38)% for lesions in or near MR susceptibility artifact regions not corrected for segmentation errors compared to CT-based attenuation correction taken as a reference. The SUV underestimation of PETorig in the B1 and B2 regions is caused by the underestimation of the attenuation leading to an undercorrection in the reconstruction. Keeping in mind that the soft-tissue filling still underestimates the attenuation in the metal implant

2723

G Schramm et al

Phys. Med. Biol. 59 (2014) 2713

region, the true SUV underestimations in B1 and B2 regions can be expected to be even higher (about 15%). In contrast, in L2 regions, where the lung is filled with the soft-tissue attenuation coefficient, PETorig clearly overestimates the SUV (120 ± 47%). In L1 regions, the situation is more complicated since part of the lung is assigned with the air and part with the soft-tissue attenuation coefficient. Thus, regions with overestimation and regions with underestimation of the attenuation in the lung exist (see figure 3(d)). On an average, PETorig overestimates the reconstructed SUVs in six L1 patients whereas it underestimated the SUVs in one patient only. The calculated average SUV difference in the lung varies strongly from an underestimation of −2% to an overestimation of 48%. In all patients having metal implants in the sternum, the sternum was wrongly filled with the lung tissue attenuation coefficient in MRMapcor . This wrong assignment occurs because the signal void in the sternum is closed from the anterior side. Consequently, as shown in figure 3 (d), the signal void of the sternum stays connected to the low signal region of the lung in the atMR scan and is thus assigned with the lung tissue attenuation coefficient. However, since the volume of the sternum is relatively small compared to that of the lung, it can be expected that the influence of the wrong attenuation coefficient in the sternum in MRMapcor is much smaller than the influence of wrong attenuation coefficients in the complete lung in MRMaporig . A favorable feature of the developed algorithm is its ability to recover truncated regions caused by the limited FOV of the MR scanner (Delso et al 2010, Schramm et al 2013b). Thus an additional truncation compensation is not necessary. Considering the quantitative influence of truncation artifacts on the reconstructed SUVs inside the body, it could be shown in Schramm et al (2013b) that this effect is much smaller than the SUV differences caused by metal-implant-induced segmentation errors found in this study. So far, we have applied our algorithm only to patients examined with [18 F]-FDG and [18 F]NaF because up to now no patients with metal implants have been examined with other tracers at our site. We are confident that our algorithm will work with other tracers as well, because it performs well in [18 F]-NaF patients, where soft tissue uptake is very low and PET-based patient contour delineation is more difficult. However, further testing has to be performed to validate this assumption. Concerning the determination of the patient’s outer contour we use the same method to calculate the PET background intensity threshold as described in Schramm et al (2013b) where it was shown that this approach recovers the extent of the arms correctly. The PET-derived estimation of the outer contour in the metal artifact regions worked for all patients except for patients 2 and 3 where the extent of the knees was underestimated by approximately one voxel layer. This underestimation is caused by the very weak tracer uptake in the knees in PETorig which is a limitation of the PET-based contour delineation. Of course, this limitation is most pronounced for huge PET signal voids near the surface caused by metal implants. Another limitation of the algorithm is the fact that cavities that truly contain air (e.g. the trachea) are filled with the soft-tissue attenuation coefficient as well. In order to estimate the influence of this effect we have performed an additional reconstruction where we used an MRMap with the correct air attenuation coefficient in the manually segmented trachea. The resulting SUV difference ranges from −1% to +7% with a mean of +1%. An alternative approach to achieve metal artifact free attenuation maps is the simultaneous reconstruction of PET emission and attenuation data as described in Nuyts et al (1999), Salomon et al (2011), Rezaei et al (2012), Nuyts et al (2013). An application of these algorithms to patients with metal implants would be very interesting, to see whether the correct attenuation coefficients of the implants could be recovered. Moreover, recently it was

2724

G Schramm et al

Phys. Med. Biol. 59 (2014) 2713

shown in Bezrukov et al (2013) that a combination of atlas-based and segmentation-based MRAC is also capable to compensate for metal-implant-induced artifacts. With respect to the magnitude of the observed SUV differences caused by metal-implantinduced segmentation errors in MRMaporig , care should be taken in the assessment of therapy response based on SUVs before and after surgery. For patients with metal implants inside the body who were examined in PET/MR, the MRMap needs to be visually inspected thoroughly. In the presence of segmentation errors, the MRMap has to be corrected and the emission data should be reconstructed again, in order to allow a quantitative analysis in metal implant regions. The specific SUV deviation of a lesion in those regions, of course, depends on the size of the volume that contains wrong attenuation coefficients and on the distance between the lesion and that region. As our analysis shows, deviations up to −62% have to be expected. 5. Conclusion Metal-implant-induced signal voids lead to severe segmentation errors in MR-based attenuation maps directly affecting the quantitative accuracy of the reconstructed PET data. The proposed algorithm for compensation of these artifacts and the associated segmentation errors works fully automatic and is able to correct all metal-implant-induced segmentation errors in the MR-based attenuation maps of 13 patients by filling of these regions with the soft-tissue or lung tissue attenuation coefficient as suggested in Ladefoged et al (2013). The algorithm improves PET/MR quantification distinctly and is suitable for use in clinical routine.

References Berger M, Hubbell J, Seltzer S, Chang J, Coursey J, Sukumar R, Zucker D and Olsen K 2010 XCOM: Photon Cross Section Database (version 1.5) (Gaithersburg, MD: National Institute of Standards and Technology) (http://physics.nist.gov/xcom) Bezrukov I, Schmidt H, Mantlik F, Schwenzer N, Brendle C, Sch¨olkopf B and Pichler B 2013 MR-based attenuation correction methods for improved PET quantification in lesions within bone and susceptibility artifact regions J. Nucl. Med. 54 1768–74 Bockisch A, Beyer T, Antoch G, Freudenberg L S, K¨uhl H, Debatin J F and M¨uller S P 2004 Positron emission tomography/computed tomography–imaging protocols, artifacts, and pitfalls Mol. Imaging Biol. 6 188–99 Burger C, Goerres G, Schoenes S, Buck A, Lonn A H R and Von Schulthess G K 2002 PET attenuation coefficients from CT images: experimental evaluation of the transformation of CT into PET 511-keV attenuation coefficients Eur. J. Nucl. Med. Mol. Imaging 29 922–7 Delso G, F¨urst S, Jakoby B, Ladebeck R, Ganter C, Nekolla S G, Schwaiger M and Ziegler S I 2011 Performance measurements of the Siemens mMR integrated whole-body PET/MR scanner J. Nucl. Med. 52 1914–22 Delso G, Martinez-M¨oller A, Bundschuh R A, Nekolla S G and Ziegler S I 2010 The effect of limited MR field of view in MR/PET attenuation correction Med. Phys. 37 2804–12 Goerres G, Hany T, Kamel E, von Schulthess G and Buck A 2002 Head and neck imaging with PET and PET/CT: artefacts from dental metallic implants Eur. J. Nucl. Med. Mol. Imaging 29 367–70 Goerres G, Ziegler S and Burger C 2003 Radiology artifacts at PET and PET/CT caused by metallic hip prosthetic material Radiology 226 577–84 Hofheinz F, P¨otzsch C, Oehme L, Beuthien-Baumann B, Steinbach J, Kotzerke J and van den Hoff J 2012 Automatic volume delineation in oncological PET. Evaluation of a dedicated software tool and comparison with manual delineation in clinical data sets Nuklearmedizin. 51 9–16 Hofmann M, Pichler B, Sch¨olkopf B and Beyer T 2009 Towards quantitative PET/MRI: a review of MR-based attenuation correction techniques Eur. J. Nucl. Med. Mol. Imaging 36 (Suppl. 1) S93–104 Hu Z et al 2009 MR-based Attenuation correction for a whole-body sequential PET/MR system NSS/MIC’09: IEEE Nuclear Science Symp. Conf. Record pp 3508–12 2725

G Schramm et al

Phys. Med. Biol. 59 (2014) 2713

Keereman V, Holen R V, Mollet P and Vandenberghe S 2011 The effect of errors in segmented attenuation maps on PET quantification Med. Phys. 38 6010 Keller S H, Holm S, Hansen A E, Sattler B, Andersen F, Klausen T L, Højgaard L, Kjær A and Beyer T 2013 Image artifacts from MR-based attenuation correction in clinical, whole-body PET/MRI Magn. Reson. Mater. Phys. Biol. Med. 26 173–81 Ladefoged C N, Andersen F L, Keller S H, L¨ofgren J, Hansen A E, Holm S, Højgaard L and Beyer T 2013 PET/MR imaging of the pelvis in the presence of endoprostheses: reducing image artifacts and increasing accuracy through inpainting Eur. J. Nucl. Med. Mol. Imaging 40 594–601 Martinez-M¨oller A, Souvatzoglou M, Delso G, Bundschuh R A, Chefd’hotel C, Ziegler S I, Navab N, Schwaiger M and Nekolla S G 2009 Tissue classification as a potential approach for attenuation correction in whole-body PET/MRI: evaluation with PET/CT data J. Nucl. Med. 50 520–6 Nuyts J, Bal G, Kehren F, Fenchel M, Michel C and Watson C 2013 Completion of a truncated attenuation image from the attenuated PET emission data IEEE Trans. Med. Imaging 32 237–46 Nuyts J, Dupont P, Stroobants S, Benninck R, Mortelmans L and Suetens P 1999 Simultaneous maximum a posteriori reconstruction of attenuation and activity distributions from emission sinograms IEEE Trans. Med. Imaging 18 393–403 Rezaei A, Defrise M, Bal G, Michel C, Conti M, Watson C and Nuyts J 2012 Simultaneous reconstruction of activity and attenuation in time-of-flight PET IEEE Trans. Med. Imaging 31 2224–33 Salomon A, Goedicke A, Schweizer B, Aach T and Schulz V 2011 Simultaneous reconstruction of activity and attenuation for PET/MR IEEE Trans. Med. Imaging 30 804–13 Schenk J F 1996 The role of magnetic susceptibility in magnetic resonance imaging: MRI magnetic compatibility of the first and second kinds Med. Phys. 23 815–50 Schramm G, Langner J, Hofheinz F, Petr J, Beuthien-Baumann B, Platzek I, Steinbach J, Kotzerke J and van den Hoff J 2013a Quantitative accuracy of attenuation correction in the Philips Ingenuity TF whole-body PET/MR system: a direct comparison with transmission-based attenuation correction Magn. Reson. Mater. Phys. Biol. Med. 26 115–26 Schramm G, Langner J, Hofheinz F, Petr J, Lougovski A, Beuthien-Baumann B, Platzek I and van den Hoff J 2013b Influence and compensation of truncation artifacts in MR-based attenuation correction in PET/MR IEEE Trans. Med. Imaging 32 2056–63 Schulz V et al 2011 Automatic, three-segment, MR-based attenuation correction for whole-body PET/MR data Eur. J. Nucl. Med. Mol. Imaging 38 138–52 Sureshbabu W and Mawlawi O 2005 PET/CT imaging artifacts J. Nucl. Med. Technol. 33 156–61 Zaidi H, Ojha N, Morich M, Griesmer J, Hu Z, Maniawski P, Ratib O, Izquierdo-Garcia D, Fayad Z A and Shao L 2011 Design and performance evaluation of a whole-body Ingenuity TF PET–MRI system Phys. Med. Biol. 56 3091–106

2726