- Email: [email protected]
Brain PET imaging optimization with time of flight and point spread function modelling
ARTICLE IN PRESS Physica Medica ■■ (2015) ■■–■■
Contents lists available at ScienceDirect
Physica Medica j o u r n a l h o m e p a g e : h t t p : / / w w w. p h y s i c a m e d i c a . c o m
Original Paper
Brain PET imaging optimization with time of flight and point spread function modelling Elena Prieto a, Josep M. Martí-Climent a,*, Verónica Morán a, Lidia Sancho b, Benigno Barbés a, Javier Arbizu b, Jose A. Richter b a b
Medical Physics Department, Clínica Universidad de Navarra, Av. Pío XII, 36. 31008 Pamplona, Spain Nuclear Medicine Department, Clínica Universidad de Navarra, Av. Pío XII, 36. 31008 Pamplona, Spain
A R T I C L E
I N F O
Article history: Received 14 May 2015 Received in revised form 29 June 2015 Accepted 1 July 2015 Available online Keywords: PET Brain Phantoms Optimization
A B S T R A C T
Purpose: To assess the influence of reconstruction algorithms and parameters on the PET image quality of brain phantoms in order to optimize reconstruction for clinical PET brain studies in a new generation PET/CT. Methods: The 3D Hoffman phantom that simulates 18F-fluorodeoxyglucose (FDG) distribution was imaged in a Siemens Biograph mCT TrueV PET/CT with Time of Flight (TOF) and Point Spread Function (PSF) modelling. Contrast-to-Noise Ratio (CNR), contrast and noise were studied for different reconstruction models: OSEM, OSEM + TOF, OSEM + PSF and OSEM + PSF + TOF. The 2D multi-compartment Hoffman phantom was filled to simulate 4 different tracers’ spatial distribution: FDG, 11C-flumazenil (FMZ), 11C-Methionine (MET) and 6-18F-fluoro-l-dopa (FDOPA). The best algorithm for each tracer was selected by visual inspection. The maximization of CNR determined the optimal parameters for each reconstruction. Results: In the 3D Hoffman phantom, both noise and contrast increased with increasing number of iterations and decreased with increasing FWHM. OSEM + PSF + TOF reconstruction was generally superior to other reconstruction models. Visual analysis of the 2D Hoffman brain phantom suggested that OSEM + PSF + TOF is the optimum algorithm for tracers with focal uptake, such as MET or FDOPA, and OSEM + TOF for tracers with diffuse cortical uptake (i.e. FDG and FMZ). Optimization of CNR demonstrated that OSEM + TOF reconstruction must be performed with 2 iterations and a filter FWHM of 3 mm, and OSEM + PSF + TOF reconstruction with 4 iterations and 1 mm FWHM filter. Conclusions: Optimization of reconstruction algorithm and parameters has been performed to take particular advantage of the last generation PET scanner, recommending specific settings for different brain PET radiotracers. © 2015 Associazione Italiana di Fisica Medica. Published by Elsevier Ltd. All rights reserved.
Introduction Positron Emission Tomography (PET) reconstruction has advanced considerably in the last years, achieving a dramatic improvement in the quality of clinical PET images. Particularly, two important components of the physics modelling of the imaging process have been included in the reconstruction: detector point spread function (PSF) and time of flight (TOF) [1–3]. PSF algorithms consist of measuring or simulating the point spread function at several points within the field of view of the PET scanner to incorporate this information into the iterative
* Corresponding author. Medical Physics Department, Clínica Universidad de Navarra, Av. Pío XII, 36. 31008 Pamplona, Spain. Tel.: +0034 948 255400; fax: +0034 948 296500. E-mail address: [email protected] (J.M. Martí-Climent).
reconstruction process. This technique has been shown to improve spatial resolution, contrast-to-noise ratio (CNR) and contrast [2], and subsequently lesion detectability [4]. However, PSF-based reconstruction can cause edge artefacts, which appear as overshoot at sharp intensity transitions [5]. Time of flight techniques consist of measuring the arrival-time difference between the two coincident photons to improve localization of the annihilation event along the line of response [3]. It has been proven that when this information is incorporated into the reconstruction process, the algorithm converges with fewer iterations and CNR on the final image improves [6]. The expected gain in CNR is related to the ratio between the diameter of the positronemitting distribution and the timing resolution of the tomograph [3]. Although the benefit of these techniques with respect to the classic iterative reconstruction has been widely studied, optimization of the reconstruction parameters is necessary to obtain the best
http://dx.doi.org/10.1016/j.ejmp.2015.07.001 1120-1797/© 2015 Associazione Italiana di Fisica Medica. Published by Elsevier Ltd. All rights reserved.
Please cite this article in press as: Elena Prieto, et al., Brain PET imaging optimization with time of flight and point spread function modelling, Physica Medica (2015), doi: 10.1016/ j.ejmp.2015.07.001
ARTICLE IN PRESS E. Prieto et al./Physica Medica ■■ (2015) ■■–■■
2
performance for these algorithms. This optimization has already been investigated for oncological PET [7–9]. Nevertheless, few investigations have focused in the field of brain PET [10] and a standard reconstruction protocol has not been established yet. In fact, recent brain studies performed with last generation PET scanners are using classical OSEM algorithm and do not take advantage of the advanced reconstruction techniques maybe due to the lack of an optimization study [11–13]. Within this framework, the aim of this study is to investigate the influence of reconstruction parameters on PET image quality using brain phantoms as well as to select the best reconstruction algorithm and the optimal reconstruction parameters for clinical PET brain studies with different tracers. Methods PET tomograph PET images were acquired on a Siemens Biograph mCT PET/CT scanner (Siemens Medical Solutions, Hoffman Estates, IL) with an extended axial field of view (21.8 cm), PSF depth dependent resolution recovery and TOF acquisition and reconstruction. The PET system consists of 4 rings of LSO blocks and it is integrated with a 64-slice spiral CT scanner. The performance characterization of this PET/CT system according to National Electrical Manufacturers Association (NEMA) protocols can be found elsewhere [14]. Image quality as a function of reconstruction algorithm and parameters 3D Hoffman phantom preparation and reconstruction The 3D Hoffman Brain Phantom (Data Spectrum Corporation, Hillsborough, NC) consists of several polymethylacrylate plates which simulate cross sections of the brain stacked into a cylindrical tank (diameter 18 cm, height 12 cm) [15]. This phantom represents the radioisotope distribution as it would appear in a 18 Ffluorodeoxyglucose (FDG) PET brain study with a 4:1 uptake ratio between grey and white matter. The 3D Hoffman phantom was filled with a radioactive solution of FDG. The total activity at the start of the scan was 37 MBq, representing the activity concentration of a clinical FDG study after 40 minutes of an uptake period of 370 MBq [16]. The protocol consisted of a CT scan followed by a list-mode emission scan of 10 minutes (standard protocol for a clinical brain PET). Four different reconstruction protocols were used to process the sinogram data: baseline ordered-subsets expectation maximization (OSEM) algorithm, OSEM reconstruction with the PSF model, OSEM with TOF information and OSEM with both PSF and TOF models. For each reconstruction several parameters have to be defined. In this study, some of these parameters are considered fixed and are maintained constant through the whole study. Particularly, 24 subsets (or 21 subsets for OSEM + TOF and
OSEM + PSF + TOF algorithms), an image matrix of 400 × 400 and a zoom factor of 2 were used (pixel size 1.0 × 1.0 mm). The full width at half maximum (FWHM) of the post-smoothing Gaussian filter and the number of iterations are considered variable, and their optimization is the scope of this study. Particularly, image quality was studied as a function of iteration number (from 1 to 10) with the Gaussian filter FWHM fixed to 2 mm, and then as a function of filter width (FWHM from 1 to 10 mm) with the number of iterations fixed to 2 iterations. This evaluation was carried out for the four reconstruction algorithms. Data analysis In order to evaluate reconstruction algorithm performance, image quality was measured through quantitative figures of merit. With this aim, irregular volumes of interest (VOI) were drawn over the CT images, using the PMOD software (PMOD Technologies Ltd., Adliswil, Switzerland). Particularly, VOIs over cortical grey matter (GM), putamen (PT) and caudate (CD) were drawn using the isocontour tool covering at least 9 CT planes (slice thickness 1.5 mm) with final volumes of 25.3 cm3, 3.2 cm3 and 2.0 cm3 respectively. Additionally, a white matter (WM) region was manually drawn (2.4 cm3) (Fig. 1). The VOIs were then transferred onto the PET images to compute the mean (μ) and standard deviation (σ) of the activity concentration in each VOI. Quantitative evaluation was based on three metrics: contrast recovery, noise and contrast-to-noise ratio. The contrast recovery (CR) in each VOI was defined as the ratio between the measured contrast and the expected contrast, defining the contrast as the ratio between the mean (μ) concentration in the volume of interest (GM, CD or PT) and the mean concentration in white matter:
CRVOI
μVOI μWM (% ) = 100 ⋅ 4
The noise was computed as the ratio between the standard deviation and the average concentration in the reference region of white matter. This definition of noise refers to the figure of merit ‘image roughness’ as defined by Tong et al. [4]:
Noise ( % ) = 100 ⋅
σ WM μWM
Contrast-to-noise ratio (CNR) was computed as the difference between the signal, defined as the mean concentration in the structure of interest (GM, CD or PT), and the background, defined as the mean concentration in the white matter, compared to the noise (σ) in the background:
CNRVOI =
μVOI − μWM σ WM
Figure 1. Three representative slices of the volumes of interest drawn over the CT image of the 3D Hoffman brain phantom: cortical grey matter (blue), caudate (green), putamen (pink), and white matter region (red). (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)
Please cite this article in press as: Elena Prieto, et al., Brain PET imaging optimization with time of flight and point spread function modelling, Physica Medica (2015), doi: 10.1016/ j.ejmp.2015.07.001
ARTICLE IN PRESS E. Prieto et al./Physica Medica ■■ (2015) ■■–■■
The maximum CNR in the volumes of interest determined the optimal reconstruction parameters for clinical studies. Optimization for different radiotracers Due to the fact that the 3D Hoffman phantom can only simulate the activity distribution of FDG (i.e. high uptake in the whole grey matter), the Hoffman multi-compartment 2D brain phantom (Data Spectrum Corporation, Hillsborough, NC), which contains 7 grey matter compartments that may be separately filled, was used. In this study, 4 different tracers were simulated: FDG, 11C-flumazenil (FMZ), 6-18F-fluoro-l-dopa (FDOPA) and 11C-Methionine (MET). A normal brain distribution was simulated for FDG, FMZ and FDOPA and a brain tumor for MET. Despite the fact that typical concentration levels in the brain might be different for each tracer [17], the 2D Hoffman phantom was filled with the same FDG concentration in the principal compartment of the phantom as in the 3D phantom (30.8 kBq/cc) and with a ratio of 4:1 between the hot and background areas. The protocol for each simulated tracer consisted of a CT scan followed by a list-mode PET scan of 10 minutes. Images were reconstructed with the four algorithms and with standard reconstruction parameters (2 iterations, 21 or 24 subsets and a filter width of 2 mm). As recommended in the phantom user’s manual, all reconstructed slices were added together to form one thick slice. Analysis of these PET images was performed visually by two experienced observers (one nuclear medicine physician with 20 years
3
of experience and one physicist with 8 years of experience) to select the optimum reconstruction algorithm for each tracer. Once the optimal reconstruction algorithm is selected for a given tracer, the optimal reconstruction parameters must be selected for that algorithm. In order to perform a realistic quantitative analysis, the CNR with varying reconstruction parameters was measured in the 3D Hoffman brain phantom. The optimization was conducted with the previously described quantification procedure based on CNR maximization, varying the number of iterations from 1 to 5 and the filter width from 1 to 5 mm in 1 mm increments. For MET and FDOPA, CNR was optimized using only the small VOIs (caudate and putamen), while for FDG and FMZ the three VOIs (caudate, putamen and grey matter) were used. Results CNR, contrast and noise as function of reconstruction parameters PET images of the 3D Hoffman brain phantom with the 4 reconstruction algorithms and different number of iterations, together with the image quality metrics plotted as a function of the number of iterations, are presented in Fig. 2. In the graph, CNR and CR are presented only for the grey matter volume; the other VOIs (caudate and putamen) presented the same tendencies with lower CNR and contrast recovery values due to their smaller volume. From visual inspection, all the structures of the 3D Hoffman brain phantom could
Figure 2. A representative slice of the 3D Hoffman brain phantom reconstructed with different algorithms and different numbers of iterations. Trend of CR in grey matter, noise and CNR in grey matter are plotted as a function of the iteration number for the 4 reconstruction algorithms. The filter size is fixed to 2 mm.
Please cite this article in press as: Elena Prieto, et al., Brain PET imaging optimization with time of flight and point spread function modelling, Physica Medica (2015), doi: 10.1016/ j.ejmp.2015.07.001
ARTICLE IN PRESS 4
E. Prieto et al./Physica Medica ■■ (2015) ■■–■■
be clearly recognized in the PET image, independently of the reconstruction algorithm and the number of iterations. Regarding quantitative metrics, contrast recovery increased with the iteration number eventually reaching a plateau. The incorporation of PSF produced significantly higher contrasts. In fact, CR values converge to 69% without PSF and to 78% with PSF. The nonTOF reconstructions converge to slightly lower contrast values than their corresponding TOF image, but CR differences are lower than 3%. Contrast converged faster when TOF information is included in the reconstruction. The most accurate contrast values were obtained for the OSEM + PSF + TOF reconstruction. Noise monotonically increased with iteration number for all the reconstruction algorithms. Values varied from around 10% for 1 iteration to values greater than 22% for 10 iterations and PSF reconstructions, and greater than 25% for 10 iterations and nonPSF algorithms. It is noteworthy that PSF models achieved a reduction in the noise level of approximately 2–5% with respect to the OSEM or OSEM + TOF reconstructions. In most cases, reconstructions that showed higher levels of contrast also had higher noise levels. CNR shows an important variability with the number of iterations, decreasing monotonically for more than 2 iterations. In general, for matched reconstruction parameters, both PSF and TOF models improved CNR with respect to the baseline OSEM reconstruction. However, CNR improvement with PSF is greater than 40% for 5 or more iterations, while improvement with TOF is much lower, around 2–6% in most cases. Analogously, Fig. 3 presents the PET images and the quantitative metrics for the 4 algorithms and different filter widths. Images
became blurred as the filter width increased. Regarding the quantitative analysis, metrics are presented only for grey matter. Caudate and putamen presented the same tendencies for CNR and contrast but with lower values. The analysis of the quantitative metrics shows that contrast recovery monotonically decreased with increasing FWHM, getting away from the true contrast value. OSEM + PSF + TOF reconstruction obtained the best CR for all filter widths with a CR value of 73% for a 1 mm filter and 50% for a 10 mm filter. On the other hand, the noise has a trend to decrease with increasing filter FWHM. The OSEM and OSEM + TOF reconstructions are the most influenced by the Gaussian filter, with a reduction in the noise level of 27% and 26% respectively when filter width is increased from 1 to 4 mm. For the reconstructions with PSF, with or without TOF, noise level only varied at 5.3% between the largest and smallest filters. At matched iterations, noise level is higher when TOF is introduced in both the OSEM and OSEM + PSF reconstructions. In fact, for wide filters, OSEM and OSEM + PSF reconstructions present a noise level of 12% while OSEM + TOF and OSEM + PSF + TOF suffer higher noise level (14%). Figure 3 shows that the different reconstruction algorithms present clearly different CNRs when the size of the filter is small (less than 4 mm) but differences tend to disappear for wider filters. The tendency of CNR is monotonically decreasing when PSF is included in the reconstruction, while without PSF, CNR increases with the filter width until an FWHM of 3 mm, and then decreases.
Figure 3. A representative slice of the 3D Hoffman brain phantom reconstructed with different algorithms and different filter width. Trend of CR in grey matter, noise and CNR in grey matter plotted as a function of the filter width for the 4 reconstruction algorithms. The iteration number is fixed to 2.
Please cite this article in press as: Elena Prieto, et al., Brain PET imaging optimization with time of flight and point spread function modelling, Physica Medica (2015), doi: 10.1016/ j.ejmp.2015.07.001
ARTICLE IN PRESS E. Prieto et al./Physica Medica ■■ (2015) ■■–■■
Optimization for different radiotracers Figure 4 represents the 2D Hoffman phantom simulating 4 different tracer distributions in the brain and reconstructed with the four reconstruction algorithms. By visual inspection, introduction of TOF does not have great impact on the image quality. However, TOF had demonstrated an improvement in CNR in the previous quantitative analysis over the 3D Hoffman phantom (Figs. 2 and 3). Therefore, qualitative analysis is not influenced by TOF but quantitative analysis suggests that TOF should always be included in the reconstruction. Regarding the introduction of PSF modelling, an edge enhancement between structures with different concentrations can be appreciated both in the PET images and especially in the activity profiles. This effect is especially pronounced between grey matter (with 4 times the activity concentration of white matter) and the background (with 0 activity concentration) due to the sharpness of the transition, while it is less marked between the inner structures of the
5
brain (transitions from grey matter to white matter with a concentration ratio of 4:1). The influence of this effect should be discussed separately according to the uptake pattern of the different brain PET tracers. FDOPA and MET belong to a group of tracers with focal uptake within the brain, in small structures such as striatum for FDOPA or in a brain tumor for MET. In these cases, the ring effect between the areas of interest and the unspecific uptake areas could be very useful for detecting striatal areas with reduced uptake of FDOPA and small foci with MET. On the other hand, FDG and FMZ uptake is distributed on the whole cortex and the uptake overestimation in the edge is significant. The diagnosis for these tracers is based on detecting cortical areas with higher or lower uptake than the normal cortex. In these cases, the interpretation can be biased due to the ringing artefact. As a consequence, PSF should be avoided for tracers such as FDG of FMZ. Thus, the optimal reconstruction algorithm for these tracers would be OSEM + TOF, while for the rest of tracers OSEM + PSF + TOF can be used. Therefore, these are the two algorithms that required optimization of reconstruction parameters.
Figure 4. 2D Hoffman brain phantom filled to simulate different tracer’s distributions. Images were reconstructed with OSEM, OSEM + TOF, OSEM + PSF and OSEM + PSF + TOF. Reconstruction parameters were fixed to 2 iterations and a 2 mm filter FWHM. The sum of the slices of the 3D PET image is presented. Profile curves were obtained across the dotted line.
Please cite this article in press as: Elena Prieto, et al., Brain PET imaging optimization with time of flight and point spread function modelling, Physica Medica (2015), doi: 10.1016/ j.ejmp.2015.07.001
ARTICLE IN PRESS E. Prieto et al./Physica Medica ■■ (2015) ■■–■■
6
FDG AND FMZ OPTIMIZATION
MET AND FDOPA OPTIMIZATION
OSEM+TOF VOIs: CD,PT,GM
OSEM+PSF+TOF VOIs: CD,PT
CNRaverage
9 8
8
7
7
6
6
5
CRaverage
9
1
2
3
4
5
5
70
70
60
60
50
50
40
1
2
3
Iteration number 1 mm
4
5
2 mm
40
3 mm
1
2
1
2
3
4
5
3
4
5
Iteration number
4 mm
5 mm
Figure 5. Optimization of the number of iterations and the filter width for maximization of CNR in two reconstruction algorithms: OSEM + TOF and OSEM + PSF + TOF. Data are obtained for the VOIS represented in Fig. 1 applied to the 3D Hoffman phantom.
Once the algorithm is selected for each tracer, the combination of parameters that yields to the best image quality must be decided. As described in the Methods section, this process was based on the 3D Hoffman phantom and on the analysis of CNR. Figure 5 shows the CNR values, averaging caudate and putamen (the small VOIs) for MET and FDOPA and averaging the three VOIs (CD, PT, GM) for FDG and FMZ optimization. The highest CNR for OSEM + TOF reconstruction was obtained for 2 iterations and a filter FWHM of 3 mm (average CNR = 8.8). For OSEM + PSF + TOF reconstruction, the optimal parameters were 4 iterations and 1 mm FWHM filter (average CNR = 8.9), although almost the same CNR value is obtained for 3 iterations and 1 mm FWHM filter (average CNR = 8.8). CR values for all the combinations of parameters considered in this analysis are also presented in Fig. 5. It is noted that the conditions for maximum CNR do not provide maximum CR values. In fact, maximum CR values are obtained for the maximum number of iterations and the minimum filter width. Discussion In this study the influence of reconstruction parameters on the image quality of brain PET studies has been evaluated in a new generation PET/CT scanner. Although Nagaki et al. [10] have previously studied the influence of the advanced reconstruction techniques for brain PET, this is the first study to perform a systematic optimization of reconstruction parameters for each algorithm. In fact, a standard reconstruction for brain PET in last generation PET tomographs has not been established yet and recent brain studies performed with modern scanners reconstruct images with classical OSEM algorithm and do not take advantage of the available advanced reconstruction protocols [11–13]. It has been demonstrated that the optimal reconstruction algorithm for tracers with global cortical uptake, such as FDG, would be iterative OSEM with TOF but without PSF to avoid overshooting
in the edges of the cortex. On the other hand, for tracers with focal uptake (i.e. FDOPA and MET or labelled amino acids) the reconstruction with both techniques, PSF and TOF, should be used. In fact, this algorithm has been proven to provide the best image quality metrics. Besides, reconstruction parameters have been studied to optimize algorithm performance. The OSEM + PSF + TOF algorithm has been demonstrated to require 4 (or 3) iterations and a 1 mm filter for maximal performance. In agreement with our results, Kadrmas et al. [18] have previously demonstrated that the PSF algorithm achieves maximal lesion-detection with little or no filtering in a whole body phantom. On the other hand, for the OSEM + TOF algorithm the optimal parameters were 2 iterations and a filter of 3 mm. In the literature, optimization studies for this same PET/CT but different clinical applications have been already published (i.e. whole body FDG [9] and 90Y PET [19]), and it should be remarked that different reconstruction parameters are optimum for different clinical situations. An important novelty of our study is that we have performed the optimization specifically for brain studies, taking into account the different types of radiotracers that can be used in the clinical setting. The influence of TOF and/or PSF algorithms in the image quality has been already studied in the literature, but most analyses are based on the NEMA phantom with spheres simulating oncological PET. Despite this, our findings are consistent with previous published results. Time of flight has been previously demonstrated to converge faster and to obtain an important CNR improvement, especially for large diameter objects [20]. In fact, Lois et al. [7] demonstrated that the CNR improvement is a factor of 2.1 for a 40 cm object and a timing resolution of 600 ps. Nagaki et al. [10] demonstrated in a series of clinical brain studies that the addition of TOF resulted in an improvement of 23% for the signal to noise ratio, but differences in resolution or contrast were not statistically significant. Our study with the 3D Hoffman brain phantom has confirmed the
Please cite this article in press as: Elena Prieto, et al., Brain PET imaging optimization with time of flight and point spread function modelling, Physica Medica (2015), doi: 10.1016/ j.ejmp.2015.07.001
ARTICLE IN PRESS E. Prieto et al./Physica Medica ■■ (2015) ■■–■■
acceleration of convergence with TOF and also a slight improvement in CNR. However, this improvement is not very pronounced due to the relatively small diameter that the phantom and the human head have. Regarding PSF reconstruction, our study has demonstrated a clear image quality improvement in brain PET, in accordance with previous works that have reported a significant noise level reduction in oncologic studies [4]. According to Nagaki et al. [10], PSF achieves an improvement in resolution and signal to noise ratio while differences in contrast were not significant in clinical brain PET. The ringing artefact produced with this reconstruction has been widely mentioned in the literature [21]. However, this is the first time that this effect has been evaluated for different brain PET radiotracers and that its potential impact in clinical diagnosis has been questioned. In fact, based on our analysis, the use of PSF should be avoided in the clinical setting for tracers such as FDG or FMZ with high uptake in grey matter. Previous studies with the Hoffman phantom were based on visual analysis of the image or on visual analysis of the line profiles through a reconstructed slice [16,22]. Our procedure for quantitative analysis is similar to the one used by Bettinardi et al. [23], who, as part of the performance characterization of the PET/CT Discovery-690 scanner, included an analysis of the 3D Hoffman brain phantom. A similar procedure was described by van Velden et al. [24]. Nevertheless, their analyses were only based on the contrast ratio of a grey matter region. As an improvement, we have also considered caudate and putamen as areas of interest and included three different metrics in the quantitative analysis of the image quality. The first part of this study consisted of the evaluation of the tendency of the image quality metric for different algorithms and parameters. Although the general idea that noise increases while contrast decreases with higher number of iterations and lower filter is a well-known fact, the systematic analysis performed in our study shows the detailed tendency of both parameters and provides deeper knowledge to understand the effect of these two components in the CNR of the image. We have selected the CNR value as the most important parameter and the maximization of this metric was performed to optimize the reconstruction parameters. This approach has been used previously in the literature [9]. However, if quantitative accuracy takes priority instead of the global image quality, optimization should be performed on the basis of CR instead of CNR. In this case, as shown in Fig. 5, the reconstruction should be performed with the highest number of iterations and the narrowest filter. The CR values obtained in our study can be compared with the values proposed by the Japanese Alzheimer’s Disease Neuroimaging Initiative (J-ADNI). They proposed 55% or higher GM/WM contrast as a practical qualification criterion, which is met by most of the currently used PET cameras [25]. Even though the equation to define CR values is slightly different, this criterion is fulfilled in our study for all the reconstruction algorithms with 1–5 iterations and an FWHM from 1 to 5 mm. Using a PET/CT Discovery-690 and an analysis similar to ours, Bettinardi et al. [23] have found CR values near 87% with PSF and around 80% without PSF for 10 iterations or more. These values are slightly higher than the CR obtained in our study, maybe due to differences in the VOI definition or in the acquisition time. This study had some limitations. The most important limitation is that a single acquisition was used for all the reconstructions. Therefore, the uncertainty of the data has not been measured and statistical analysis was not feasible. The measurement of uncertainty would have required several realizations of the experiment, which might be problematic due to the time and FDG required, or the estimation through the measurement of background variability defined by NEMA, which has demonstrated a linear relationship with noise across multiple realizations [4]. However, this figure of
7
merit is not measurable in the Hoffman brain phantom due to its shape. Owing to these issues, the evaluation of reconstruction algorithms with a single experiment is frequently found in the literature [7,9,23]. However, this lack of statistical analysis means that further work is required to confirm our conclusions. Regarding other limitations, we determined the reconstruction parameters based on phantom studies but we did not evaluate image quality in clinical images. Therefore, further studies are required to validate our results in clinical patients. Also, we have only implemented and tested this methodology for a single PET scanner. However, this methodology could be reproduced on any other PET scanner with TOF and PSF capabilities. It should be noted that, for multicentre studies, the optimum image quality in each centre might not be the ultimate goal, but a homogeneous image quality and quantification. For this distinct purpose, Joshi et al. [22] have developed a specific technique based on smoothing the images from different scanner models to the minimum resolution in the considered set of tomographs. Conclusion A wide range of reconstruction algorithms and parameters have been tested in a last generation PET/CT scanner and image quality has been characterized specifically for brain studies. Both PSF and TOF have been proven to provide a substantial improvement in image quality metrics, especially in terms of contrast and contrast-tonoise ratio. However, PSF should be avoided in the reconstruction for radiotracers where the ringing artefact might hamper the clinical interpretation of the PET study. Therefore, OSEM + TOF is the optimum model for tracers with diffuse cortical uptake (i.e. FDG and FMZ), while OSEM + PSF + TOF should be used for tracers with focal uptake, such as MET or FDOPA. Optimum reconstruction parameters have been proposed for each reconstruction algorithm to take the best advantage of PET/CT in the clinical setting. Acknowledgements This work was partially funded by the Government of Spain, Institute of Health Carlos III, Ministry of Science and Innovation (project ADE 10/00028) and by Siemens Healthcare. References [1] Conti M. State of the art and challenges of time-of-flight PET. Phys Med 2009;25:1–11. [2] Panin VY, Kehren F, Michel C, Casey M. Fully 3-D PET reconstruction with system matrix derived from point source measurements. IEEE Trans Med Imaging 2006;25:907–21. [3] Budinger TF. Time-of-flight positron emission tomography: status relative to conventional PET. J Nucl Med 1983;24:73–8. [4] Tong S, Alessio A, Kinahan P. Noise and signal properties in PSF-based fully 3D PET image reconstruction: an experimental evaluation. Phys Med Biol 2010;55:1453–73. [5] Tong S, Alessio AM, Thielemans K, Stearns C, Ross S, Kinahan PE. Properties and mitigation of edge artifacts in PSF-based PET reconstruction. IEEE Trans Nucl Sci 2011;58:2264–75. [6] Surti S, Karp JS, Popescu LM, Daube-Witherspoon ME, Werner M. Investigation of time-of-flight benefit for fully 3-DPET. IEEE Trans Med Imaging 2006;25:529– 38. [7] Lois C, Jakoby BW, Long MJ, Hubner KF, Barker DW, Casey ME, et al. An assessment of the impact of incorporating time-of-flight information into clinical PET/CT imaging. J Nucl Med 2010;51:237–45. [8] Prieto E, Domínguez-Prado I, García-Velloso MJ, Peñuelas I, Richter JA, Martí-Climent JM. Impact of time-of-flight and point-spread-function in SUV quantification for oncological PET. Clin Nucl Med 2013;38:103–9. [9] Akamatsu G, Ishikawa K, Mitsumoto K, Taniguchi T, Ohya N, Baba S, et al. Improvement in PET/CT image quality with a combination of point-spread function and time-of-flight in relation to reconstruction parameters. J Nucl Med 2012;53:1716–22. [10] Nagaki A, Onoguchi M, Matsutomo N. Clinical validation of high-resolution image reconstruction algorithms in brain 18F-FDG-PET: effect of incorporating
Please cite this article in press as: Elena Prieto, et al., Brain PET imaging optimization with time of flight and point spread function modelling, Physica Medica (2015), doi: 10.1016/ j.ejmp.2015.07.001
ARTICLE IN PRESS E. Prieto et al./Physica Medica ■■ (2015) ■■–■■
8
[11]
[12]
[13]
[14]
[15] [16]
[17]
Gaussian filter, point spread function, and time-of-flight. Nucl Med Commun 2014;35:1224–32. Teune LK, Renken RJ, de Jong BM, Willemsen AT, van Osch MJ, Roerdink JB, et al. Parkinson’s disease-related perfusion and glucose metabolic brain patterns identified with PCASL-MRI and FDG-PET imaging. Neuroimage Clin 2014;5:240–4. Walter F, Cloughesy T, Walter MA, Lai A, Nghiemphu P, Wagle N, et al. Impact of 3,4-dihydroxy-6-18F-fluoro-L-phenylalanine PET/CT on managing patients with brain tumors: the referring physician’s perspective. J Nucl Med 2012;53:393–8. Lapa C, Linsenmann T, Monoranu CM, Samnick S, Buck AK, Bluemel C, et al. Comparison of the amino acid tracers 18F-FET and 18F-DOPA in high-grade glioma patients. J Nucl Med 2014;55:1611–16. Martí-Climent JM, Prieto E, Domínguez-Prado I, García-Velloso MJ, Rodríguez-Fraile M, Arbizu J, et al. Contribution of time of flight and point spread function modeling to the performance characteristics of the PET/CT Biograph mCT scanner. Rev Esp Med Nucl Imagen Mol 2013;32:13–21. Hoffman E, Cutler P, Digby W, Mazziotta J. 3-D phantom to simulate cerebral blood flow and metabolic images for PET. IEEE Trans Nucl Sci 1990;37:616–20. Baghaei H, Mawlawi O, Wang Y, Li H, Ramirez R, Kim S, et al. A comparison of five whole-body PET scanners by scanning Hoffman brain phantom. IEEE Nucl Sci Symp Conf Rec 2006;4:1973–6. Freedenberg MI, Badawi RD, Tarantal AF, Cherry SR. Performance and limitations of positron emission tomography (PET) scanners for imaging very low activity sources. Phys Med 2014;30:104–10.
[18] Kadrmas DJ, Casey ME, Conti M, Jakoby BW, Lois C, Townsend DW. Impact of time-of-flight on PET tumor detection. J Nucl Med 2009;50:1315–23. [19] Martí-Climent JM, Prieto E, Elosúa C, Rodríguez-Fraile M, Domínguez-Prado I, Vigil C, et al. PET optimization for improved assessment and accurate quantification of 90Y-microsphere biodistribution after radioembolization. Med Phys 2014;41:092503. [20] Conti M. Effect of randoms on signal-to-noise ratio in TOF PET. IEEE Trans Nucl Sci 2006;53:1188–93. [21] Snyder DL, Miller MI, Thomas LJ, Politte DG. Noise and edge artifacts in maximum-likelihood reconstructions for emission tomography. IEEE Trans Med Imaging 1987;6:228–38. [22] Joshi A, Koeppe RA, Fessler JA. Reducing between scanner differences in multi-center PET studies. Neuroimage 2009;46:154–9. [23] Bettinardi V, Presotto L, Rapisarda E, Picchio M, Gianolli L, Gilardi M. Physical performance of the new hybrid PET/CT Discovery-690. Med Phys 2011;38:5394–411. [24] van Velden FH, Kloet RW, van Berckel BN, Lammertsma AA, Boellaard R. Accuracy of 3-dimensional reconstruction algorithms for the high-resolution research tomograph. J Nucl Med 2009;50:72–80. [25] Ikari Y, Nishio T, Akamatsu G, Miya Y, Senda M. PET camera qualification criteria for brain amyloid imaging: gray/white contrast of Hoffman phantom images. J Nucl Med 2014;55(Suppl. 1):2059.
Please cite this article in press as: Elena Prieto, et al., Brain PET imaging optimization with time of flight and point spread function modelling, Physica Medica (2015), doi: 10.1016/ j.ejmp.2015.07.001
Contents lists available at ScienceDirect
Physica Medica j o u r n a l h o m e p a g e : h t t p : / / w w w. p h y s i c a m e d i c a . c o m
Original Paper
Brain PET imaging optimization with time of flight and point spread function modelling Elena Prieto a, Josep M. Martí-Climent a,*, Verónica Morán a, Lidia Sancho b, Benigno Barbés a, Javier Arbizu b, Jose A. Richter b a b
Medical Physics Department, Clínica Universidad de Navarra, Av. Pío XII, 36. 31008 Pamplona, Spain Nuclear Medicine Department, Clínica Universidad de Navarra, Av. Pío XII, 36. 31008 Pamplona, Spain
A R T I C L E
I N F O
Article history: Received 14 May 2015 Received in revised form 29 June 2015 Accepted 1 July 2015 Available online Keywords: PET Brain Phantoms Optimization
A B S T R A C T
Purpose: To assess the influence of reconstruction algorithms and parameters on the PET image quality of brain phantoms in order to optimize reconstruction for clinical PET brain studies in a new generation PET/CT. Methods: The 3D Hoffman phantom that simulates 18F-fluorodeoxyglucose (FDG) distribution was imaged in a Siemens Biograph mCT TrueV PET/CT with Time of Flight (TOF) and Point Spread Function (PSF) modelling. Contrast-to-Noise Ratio (CNR), contrast and noise were studied for different reconstruction models: OSEM, OSEM + TOF, OSEM + PSF and OSEM + PSF + TOF. The 2D multi-compartment Hoffman phantom was filled to simulate 4 different tracers’ spatial distribution: FDG, 11C-flumazenil (FMZ), 11C-Methionine (MET) and 6-18F-fluoro-l-dopa (FDOPA). The best algorithm for each tracer was selected by visual inspection. The maximization of CNR determined the optimal parameters for each reconstruction. Results: In the 3D Hoffman phantom, both noise and contrast increased with increasing number of iterations and decreased with increasing FWHM. OSEM + PSF + TOF reconstruction was generally superior to other reconstruction models. Visual analysis of the 2D Hoffman brain phantom suggested that OSEM + PSF + TOF is the optimum algorithm for tracers with focal uptake, such as MET or FDOPA, and OSEM + TOF for tracers with diffuse cortical uptake (i.e. FDG and FMZ). Optimization of CNR demonstrated that OSEM + TOF reconstruction must be performed with 2 iterations and a filter FWHM of 3 mm, and OSEM + PSF + TOF reconstruction with 4 iterations and 1 mm FWHM filter. Conclusions: Optimization of reconstruction algorithm and parameters has been performed to take particular advantage of the last generation PET scanner, recommending specific settings for different brain PET radiotracers. © 2015 Associazione Italiana di Fisica Medica. Published by Elsevier Ltd. All rights reserved.
Introduction Positron Emission Tomography (PET) reconstruction has advanced considerably in the last years, achieving a dramatic improvement in the quality of clinical PET images. Particularly, two important components of the physics modelling of the imaging process have been included in the reconstruction: detector point spread function (PSF) and time of flight (TOF) [1–3]. PSF algorithms consist of measuring or simulating the point spread function at several points within the field of view of the PET scanner to incorporate this information into the iterative
* Corresponding author. Medical Physics Department, Clínica Universidad de Navarra, Av. Pío XII, 36. 31008 Pamplona, Spain. Tel.: +0034 948 255400; fax: +0034 948 296500. E-mail address: [email protected] (J.M. Martí-Climent).
reconstruction process. This technique has been shown to improve spatial resolution, contrast-to-noise ratio (CNR) and contrast [2], and subsequently lesion detectability [4]. However, PSF-based reconstruction can cause edge artefacts, which appear as overshoot at sharp intensity transitions [5]. Time of flight techniques consist of measuring the arrival-time difference between the two coincident photons to improve localization of the annihilation event along the line of response [3]. It has been proven that when this information is incorporated into the reconstruction process, the algorithm converges with fewer iterations and CNR on the final image improves [6]. The expected gain in CNR is related to the ratio between the diameter of the positronemitting distribution and the timing resolution of the tomograph [3]. Although the benefit of these techniques with respect to the classic iterative reconstruction has been widely studied, optimization of the reconstruction parameters is necessary to obtain the best
http://dx.doi.org/10.1016/j.ejmp.2015.07.001 1120-1797/© 2015 Associazione Italiana di Fisica Medica. Published by Elsevier Ltd. All rights reserved.
Please cite this article in press as: Elena Prieto, et al., Brain PET imaging optimization with time of flight and point spread function modelling, Physica Medica (2015), doi: 10.1016/ j.ejmp.2015.07.001
ARTICLE IN PRESS E. Prieto et al./Physica Medica ■■ (2015) ■■–■■
2
performance for these algorithms. This optimization has already been investigated for oncological PET [7–9]. Nevertheless, few investigations have focused in the field of brain PET [10] and a standard reconstruction protocol has not been established yet. In fact, recent brain studies performed with last generation PET scanners are using classical OSEM algorithm and do not take advantage of the advanced reconstruction techniques maybe due to the lack of an optimization study [11–13]. Within this framework, the aim of this study is to investigate the influence of reconstruction parameters on PET image quality using brain phantoms as well as to select the best reconstruction algorithm and the optimal reconstruction parameters for clinical PET brain studies with different tracers. Methods PET tomograph PET images were acquired on a Siemens Biograph mCT PET/CT scanner (Siemens Medical Solutions, Hoffman Estates, IL) with an extended axial field of view (21.8 cm), PSF depth dependent resolution recovery and TOF acquisition and reconstruction. The PET system consists of 4 rings of LSO blocks and it is integrated with a 64-slice spiral CT scanner. The performance characterization of this PET/CT system according to National Electrical Manufacturers Association (NEMA) protocols can be found elsewhere [14]. Image quality as a function of reconstruction algorithm and parameters 3D Hoffman phantom preparation and reconstruction The 3D Hoffman Brain Phantom (Data Spectrum Corporation, Hillsborough, NC) consists of several polymethylacrylate plates which simulate cross sections of the brain stacked into a cylindrical tank (diameter 18 cm, height 12 cm) [15]. This phantom represents the radioisotope distribution as it would appear in a 18 Ffluorodeoxyglucose (FDG) PET brain study with a 4:1 uptake ratio between grey and white matter. The 3D Hoffman phantom was filled with a radioactive solution of FDG. The total activity at the start of the scan was 37 MBq, representing the activity concentration of a clinical FDG study after 40 minutes of an uptake period of 370 MBq [16]. The protocol consisted of a CT scan followed by a list-mode emission scan of 10 minutes (standard protocol for a clinical brain PET). Four different reconstruction protocols were used to process the sinogram data: baseline ordered-subsets expectation maximization (OSEM) algorithm, OSEM reconstruction with the PSF model, OSEM with TOF information and OSEM with both PSF and TOF models. For each reconstruction several parameters have to be defined. In this study, some of these parameters are considered fixed and are maintained constant through the whole study. Particularly, 24 subsets (or 21 subsets for OSEM + TOF and
OSEM + PSF + TOF algorithms), an image matrix of 400 × 400 and a zoom factor of 2 were used (pixel size 1.0 × 1.0 mm). The full width at half maximum (FWHM) of the post-smoothing Gaussian filter and the number of iterations are considered variable, and their optimization is the scope of this study. Particularly, image quality was studied as a function of iteration number (from 1 to 10) with the Gaussian filter FWHM fixed to 2 mm, and then as a function of filter width (FWHM from 1 to 10 mm) with the number of iterations fixed to 2 iterations. This evaluation was carried out for the four reconstruction algorithms. Data analysis In order to evaluate reconstruction algorithm performance, image quality was measured through quantitative figures of merit. With this aim, irregular volumes of interest (VOI) were drawn over the CT images, using the PMOD software (PMOD Technologies Ltd., Adliswil, Switzerland). Particularly, VOIs over cortical grey matter (GM), putamen (PT) and caudate (CD) were drawn using the isocontour tool covering at least 9 CT planes (slice thickness 1.5 mm) with final volumes of 25.3 cm3, 3.2 cm3 and 2.0 cm3 respectively. Additionally, a white matter (WM) region was manually drawn (2.4 cm3) (Fig. 1). The VOIs were then transferred onto the PET images to compute the mean (μ) and standard deviation (σ) of the activity concentration in each VOI. Quantitative evaluation was based on three metrics: contrast recovery, noise and contrast-to-noise ratio. The contrast recovery (CR) in each VOI was defined as the ratio between the measured contrast and the expected contrast, defining the contrast as the ratio between the mean (μ) concentration in the volume of interest (GM, CD or PT) and the mean concentration in white matter:
CRVOI
μVOI μWM (% ) = 100 ⋅ 4
The noise was computed as the ratio between the standard deviation and the average concentration in the reference region of white matter. This definition of noise refers to the figure of merit ‘image roughness’ as defined by Tong et al. [4]:
Noise ( % ) = 100 ⋅
σ WM μWM
Contrast-to-noise ratio (CNR) was computed as the difference between the signal, defined as the mean concentration in the structure of interest (GM, CD or PT), and the background, defined as the mean concentration in the white matter, compared to the noise (σ) in the background:
CNRVOI =
μVOI − μWM σ WM
Figure 1. Three representative slices of the volumes of interest drawn over the CT image of the 3D Hoffman brain phantom: cortical grey matter (blue), caudate (green), putamen (pink), and white matter region (red). (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)
Please cite this article in press as: Elena Prieto, et al., Brain PET imaging optimization with time of flight and point spread function modelling, Physica Medica (2015), doi: 10.1016/ j.ejmp.2015.07.001
ARTICLE IN PRESS E. Prieto et al./Physica Medica ■■ (2015) ■■–■■
The maximum CNR in the volumes of interest determined the optimal reconstruction parameters for clinical studies. Optimization for different radiotracers Due to the fact that the 3D Hoffman phantom can only simulate the activity distribution of FDG (i.e. high uptake in the whole grey matter), the Hoffman multi-compartment 2D brain phantom (Data Spectrum Corporation, Hillsborough, NC), which contains 7 grey matter compartments that may be separately filled, was used. In this study, 4 different tracers were simulated: FDG, 11C-flumazenil (FMZ), 6-18F-fluoro-l-dopa (FDOPA) and 11C-Methionine (MET). A normal brain distribution was simulated for FDG, FMZ and FDOPA and a brain tumor for MET. Despite the fact that typical concentration levels in the brain might be different for each tracer [17], the 2D Hoffman phantom was filled with the same FDG concentration in the principal compartment of the phantom as in the 3D phantom (30.8 kBq/cc) and with a ratio of 4:1 between the hot and background areas. The protocol for each simulated tracer consisted of a CT scan followed by a list-mode PET scan of 10 minutes. Images were reconstructed with the four algorithms and with standard reconstruction parameters (2 iterations, 21 or 24 subsets and a filter width of 2 mm). As recommended in the phantom user’s manual, all reconstructed slices were added together to form one thick slice. Analysis of these PET images was performed visually by two experienced observers (one nuclear medicine physician with 20 years
3
of experience and one physicist with 8 years of experience) to select the optimum reconstruction algorithm for each tracer. Once the optimal reconstruction algorithm is selected for a given tracer, the optimal reconstruction parameters must be selected for that algorithm. In order to perform a realistic quantitative analysis, the CNR with varying reconstruction parameters was measured in the 3D Hoffman brain phantom. The optimization was conducted with the previously described quantification procedure based on CNR maximization, varying the number of iterations from 1 to 5 and the filter width from 1 to 5 mm in 1 mm increments. For MET and FDOPA, CNR was optimized using only the small VOIs (caudate and putamen), while for FDG and FMZ the three VOIs (caudate, putamen and grey matter) were used. Results CNR, contrast and noise as function of reconstruction parameters PET images of the 3D Hoffman brain phantom with the 4 reconstruction algorithms and different number of iterations, together with the image quality metrics plotted as a function of the number of iterations, are presented in Fig. 2. In the graph, CNR and CR are presented only for the grey matter volume; the other VOIs (caudate and putamen) presented the same tendencies with lower CNR and contrast recovery values due to their smaller volume. From visual inspection, all the structures of the 3D Hoffman brain phantom could
Figure 2. A representative slice of the 3D Hoffman brain phantom reconstructed with different algorithms and different numbers of iterations. Trend of CR in grey matter, noise and CNR in grey matter are plotted as a function of the iteration number for the 4 reconstruction algorithms. The filter size is fixed to 2 mm.
Please cite this article in press as: Elena Prieto, et al., Brain PET imaging optimization with time of flight and point spread function modelling, Physica Medica (2015), doi: 10.1016/ j.ejmp.2015.07.001
ARTICLE IN PRESS 4
E. Prieto et al./Physica Medica ■■ (2015) ■■–■■
be clearly recognized in the PET image, independently of the reconstruction algorithm and the number of iterations. Regarding quantitative metrics, contrast recovery increased with the iteration number eventually reaching a plateau. The incorporation of PSF produced significantly higher contrasts. In fact, CR values converge to 69% without PSF and to 78% with PSF. The nonTOF reconstructions converge to slightly lower contrast values than their corresponding TOF image, but CR differences are lower than 3%. Contrast converged faster when TOF information is included in the reconstruction. The most accurate contrast values were obtained for the OSEM + PSF + TOF reconstruction. Noise monotonically increased with iteration number for all the reconstruction algorithms. Values varied from around 10% for 1 iteration to values greater than 22% for 10 iterations and PSF reconstructions, and greater than 25% for 10 iterations and nonPSF algorithms. It is noteworthy that PSF models achieved a reduction in the noise level of approximately 2–5% with respect to the OSEM or OSEM + TOF reconstructions. In most cases, reconstructions that showed higher levels of contrast also had higher noise levels. CNR shows an important variability with the number of iterations, decreasing monotonically for more than 2 iterations. In general, for matched reconstruction parameters, both PSF and TOF models improved CNR with respect to the baseline OSEM reconstruction. However, CNR improvement with PSF is greater than 40% for 5 or more iterations, while improvement with TOF is much lower, around 2–6% in most cases. Analogously, Fig. 3 presents the PET images and the quantitative metrics for the 4 algorithms and different filter widths. Images
became blurred as the filter width increased. Regarding the quantitative analysis, metrics are presented only for grey matter. Caudate and putamen presented the same tendencies for CNR and contrast but with lower values. The analysis of the quantitative metrics shows that contrast recovery monotonically decreased with increasing FWHM, getting away from the true contrast value. OSEM + PSF + TOF reconstruction obtained the best CR for all filter widths with a CR value of 73% for a 1 mm filter and 50% for a 10 mm filter. On the other hand, the noise has a trend to decrease with increasing filter FWHM. The OSEM and OSEM + TOF reconstructions are the most influenced by the Gaussian filter, with a reduction in the noise level of 27% and 26% respectively when filter width is increased from 1 to 4 mm. For the reconstructions with PSF, with or without TOF, noise level only varied at 5.3% between the largest and smallest filters. At matched iterations, noise level is higher when TOF is introduced in both the OSEM and OSEM + PSF reconstructions. In fact, for wide filters, OSEM and OSEM + PSF reconstructions present a noise level of 12% while OSEM + TOF and OSEM + PSF + TOF suffer higher noise level (14%). Figure 3 shows that the different reconstruction algorithms present clearly different CNRs when the size of the filter is small (less than 4 mm) but differences tend to disappear for wider filters. The tendency of CNR is monotonically decreasing when PSF is included in the reconstruction, while without PSF, CNR increases with the filter width until an FWHM of 3 mm, and then decreases.
Figure 3. A representative slice of the 3D Hoffman brain phantom reconstructed with different algorithms and different filter width. Trend of CR in grey matter, noise and CNR in grey matter plotted as a function of the filter width for the 4 reconstruction algorithms. The iteration number is fixed to 2.
Please cite this article in press as: Elena Prieto, et al., Brain PET imaging optimization with time of flight and point spread function modelling, Physica Medica (2015), doi: 10.1016/ j.ejmp.2015.07.001
ARTICLE IN PRESS E. Prieto et al./Physica Medica ■■ (2015) ■■–■■
Optimization for different radiotracers Figure 4 represents the 2D Hoffman phantom simulating 4 different tracer distributions in the brain and reconstructed with the four reconstruction algorithms. By visual inspection, introduction of TOF does not have great impact on the image quality. However, TOF had demonstrated an improvement in CNR in the previous quantitative analysis over the 3D Hoffman phantom (Figs. 2 and 3). Therefore, qualitative analysis is not influenced by TOF but quantitative analysis suggests that TOF should always be included in the reconstruction. Regarding the introduction of PSF modelling, an edge enhancement between structures with different concentrations can be appreciated both in the PET images and especially in the activity profiles. This effect is especially pronounced between grey matter (with 4 times the activity concentration of white matter) and the background (with 0 activity concentration) due to the sharpness of the transition, while it is less marked between the inner structures of the
5
brain (transitions from grey matter to white matter with a concentration ratio of 4:1). The influence of this effect should be discussed separately according to the uptake pattern of the different brain PET tracers. FDOPA and MET belong to a group of tracers with focal uptake within the brain, in small structures such as striatum for FDOPA or in a brain tumor for MET. In these cases, the ring effect between the areas of interest and the unspecific uptake areas could be very useful for detecting striatal areas with reduced uptake of FDOPA and small foci with MET. On the other hand, FDG and FMZ uptake is distributed on the whole cortex and the uptake overestimation in the edge is significant. The diagnosis for these tracers is based on detecting cortical areas with higher or lower uptake than the normal cortex. In these cases, the interpretation can be biased due to the ringing artefact. As a consequence, PSF should be avoided for tracers such as FDG of FMZ. Thus, the optimal reconstruction algorithm for these tracers would be OSEM + TOF, while for the rest of tracers OSEM + PSF + TOF can be used. Therefore, these are the two algorithms that required optimization of reconstruction parameters.
Figure 4. 2D Hoffman brain phantom filled to simulate different tracer’s distributions. Images were reconstructed with OSEM, OSEM + TOF, OSEM + PSF and OSEM + PSF + TOF. Reconstruction parameters were fixed to 2 iterations and a 2 mm filter FWHM. The sum of the slices of the 3D PET image is presented. Profile curves were obtained across the dotted line.
Please cite this article in press as: Elena Prieto, et al., Brain PET imaging optimization with time of flight and point spread function modelling, Physica Medica (2015), doi: 10.1016/ j.ejmp.2015.07.001
ARTICLE IN PRESS E. Prieto et al./Physica Medica ■■ (2015) ■■–■■
6
FDG AND FMZ OPTIMIZATION
MET AND FDOPA OPTIMIZATION
OSEM+TOF VOIs: CD,PT,GM
OSEM+PSF+TOF VOIs: CD,PT
CNRaverage
9 8
8
7
7
6
6
5
CRaverage
9
1
2
3
4
5
5
70
70
60
60
50
50
40
1
2
3
Iteration number 1 mm
4
5
2 mm
40
3 mm
1
2
1
2
3
4
5
3
4
5
Iteration number
4 mm
5 mm
Figure 5. Optimization of the number of iterations and the filter width for maximization of CNR in two reconstruction algorithms: OSEM + TOF and OSEM + PSF + TOF. Data are obtained for the VOIS represented in Fig. 1 applied to the 3D Hoffman phantom.
Once the algorithm is selected for each tracer, the combination of parameters that yields to the best image quality must be decided. As described in the Methods section, this process was based on the 3D Hoffman phantom and on the analysis of CNR. Figure 5 shows the CNR values, averaging caudate and putamen (the small VOIs) for MET and FDOPA and averaging the three VOIs (CD, PT, GM) for FDG and FMZ optimization. The highest CNR for OSEM + TOF reconstruction was obtained for 2 iterations and a filter FWHM of 3 mm (average CNR = 8.8). For OSEM + PSF + TOF reconstruction, the optimal parameters were 4 iterations and 1 mm FWHM filter (average CNR = 8.9), although almost the same CNR value is obtained for 3 iterations and 1 mm FWHM filter (average CNR = 8.8). CR values for all the combinations of parameters considered in this analysis are also presented in Fig. 5. It is noted that the conditions for maximum CNR do not provide maximum CR values. In fact, maximum CR values are obtained for the maximum number of iterations and the minimum filter width. Discussion In this study the influence of reconstruction parameters on the image quality of brain PET studies has been evaluated in a new generation PET/CT scanner. Although Nagaki et al. [10] have previously studied the influence of the advanced reconstruction techniques for brain PET, this is the first study to perform a systematic optimization of reconstruction parameters for each algorithm. In fact, a standard reconstruction for brain PET in last generation PET tomographs has not been established yet and recent brain studies performed with modern scanners reconstruct images with classical OSEM algorithm and do not take advantage of the available advanced reconstruction protocols [11–13]. It has been demonstrated that the optimal reconstruction algorithm for tracers with global cortical uptake, such as FDG, would be iterative OSEM with TOF but without PSF to avoid overshooting
in the edges of the cortex. On the other hand, for tracers with focal uptake (i.e. FDOPA and MET or labelled amino acids) the reconstruction with both techniques, PSF and TOF, should be used. In fact, this algorithm has been proven to provide the best image quality metrics. Besides, reconstruction parameters have been studied to optimize algorithm performance. The OSEM + PSF + TOF algorithm has been demonstrated to require 4 (or 3) iterations and a 1 mm filter for maximal performance. In agreement with our results, Kadrmas et al. [18] have previously demonstrated that the PSF algorithm achieves maximal lesion-detection with little or no filtering in a whole body phantom. On the other hand, for the OSEM + TOF algorithm the optimal parameters were 2 iterations and a filter of 3 mm. In the literature, optimization studies for this same PET/CT but different clinical applications have been already published (i.e. whole body FDG [9] and 90Y PET [19]), and it should be remarked that different reconstruction parameters are optimum for different clinical situations. An important novelty of our study is that we have performed the optimization specifically for brain studies, taking into account the different types of radiotracers that can be used in the clinical setting. The influence of TOF and/or PSF algorithms in the image quality has been already studied in the literature, but most analyses are based on the NEMA phantom with spheres simulating oncological PET. Despite this, our findings are consistent with previous published results. Time of flight has been previously demonstrated to converge faster and to obtain an important CNR improvement, especially for large diameter objects [20]. In fact, Lois et al. [7] demonstrated that the CNR improvement is a factor of 2.1 for a 40 cm object and a timing resolution of 600 ps. Nagaki et al. [10] demonstrated in a series of clinical brain studies that the addition of TOF resulted in an improvement of 23% for the signal to noise ratio, but differences in resolution or contrast were not statistically significant. Our study with the 3D Hoffman brain phantom has confirmed the
Please cite this article in press as: Elena Prieto, et al., Brain PET imaging optimization with time of flight and point spread function modelling, Physica Medica (2015), doi: 10.1016/ j.ejmp.2015.07.001
ARTICLE IN PRESS E. Prieto et al./Physica Medica ■■ (2015) ■■–■■
acceleration of convergence with TOF and also a slight improvement in CNR. However, this improvement is not very pronounced due to the relatively small diameter that the phantom and the human head have. Regarding PSF reconstruction, our study has demonstrated a clear image quality improvement in brain PET, in accordance with previous works that have reported a significant noise level reduction in oncologic studies [4]. According to Nagaki et al. [10], PSF achieves an improvement in resolution and signal to noise ratio while differences in contrast were not significant in clinical brain PET. The ringing artefact produced with this reconstruction has been widely mentioned in the literature [21]. However, this is the first time that this effect has been evaluated for different brain PET radiotracers and that its potential impact in clinical diagnosis has been questioned. In fact, based on our analysis, the use of PSF should be avoided in the clinical setting for tracers such as FDG or FMZ with high uptake in grey matter. Previous studies with the Hoffman phantom were based on visual analysis of the image or on visual analysis of the line profiles through a reconstructed slice [16,22]. Our procedure for quantitative analysis is similar to the one used by Bettinardi et al. [23], who, as part of the performance characterization of the PET/CT Discovery-690 scanner, included an analysis of the 3D Hoffman brain phantom. A similar procedure was described by van Velden et al. [24]. Nevertheless, their analyses were only based on the contrast ratio of a grey matter region. As an improvement, we have also considered caudate and putamen as areas of interest and included three different metrics in the quantitative analysis of the image quality. The first part of this study consisted of the evaluation of the tendency of the image quality metric for different algorithms and parameters. Although the general idea that noise increases while contrast decreases with higher number of iterations and lower filter is a well-known fact, the systematic analysis performed in our study shows the detailed tendency of both parameters and provides deeper knowledge to understand the effect of these two components in the CNR of the image. We have selected the CNR value as the most important parameter and the maximization of this metric was performed to optimize the reconstruction parameters. This approach has been used previously in the literature [9]. However, if quantitative accuracy takes priority instead of the global image quality, optimization should be performed on the basis of CR instead of CNR. In this case, as shown in Fig. 5, the reconstruction should be performed with the highest number of iterations and the narrowest filter. The CR values obtained in our study can be compared with the values proposed by the Japanese Alzheimer’s Disease Neuroimaging Initiative (J-ADNI). They proposed 55% or higher GM/WM contrast as a practical qualification criterion, which is met by most of the currently used PET cameras [25]. Even though the equation to define CR values is slightly different, this criterion is fulfilled in our study for all the reconstruction algorithms with 1–5 iterations and an FWHM from 1 to 5 mm. Using a PET/CT Discovery-690 and an analysis similar to ours, Bettinardi et al. [23] have found CR values near 87% with PSF and around 80% without PSF for 10 iterations or more. These values are slightly higher than the CR obtained in our study, maybe due to differences in the VOI definition or in the acquisition time. This study had some limitations. The most important limitation is that a single acquisition was used for all the reconstructions. Therefore, the uncertainty of the data has not been measured and statistical analysis was not feasible. The measurement of uncertainty would have required several realizations of the experiment, which might be problematic due to the time and FDG required, or the estimation through the measurement of background variability defined by NEMA, which has demonstrated a linear relationship with noise across multiple realizations [4]. However, this figure of
7
merit is not measurable in the Hoffman brain phantom due to its shape. Owing to these issues, the evaluation of reconstruction algorithms with a single experiment is frequently found in the literature [7,9,23]. However, this lack of statistical analysis means that further work is required to confirm our conclusions. Regarding other limitations, we determined the reconstruction parameters based on phantom studies but we did not evaluate image quality in clinical images. Therefore, further studies are required to validate our results in clinical patients. Also, we have only implemented and tested this methodology for a single PET scanner. However, this methodology could be reproduced on any other PET scanner with TOF and PSF capabilities. It should be noted that, for multicentre studies, the optimum image quality in each centre might not be the ultimate goal, but a homogeneous image quality and quantification. For this distinct purpose, Joshi et al. [22] have developed a specific technique based on smoothing the images from different scanner models to the minimum resolution in the considered set of tomographs. Conclusion A wide range of reconstruction algorithms and parameters have been tested in a last generation PET/CT scanner and image quality has been characterized specifically for brain studies. Both PSF and TOF have been proven to provide a substantial improvement in image quality metrics, especially in terms of contrast and contrast-tonoise ratio. However, PSF should be avoided in the reconstruction for radiotracers where the ringing artefact might hamper the clinical interpretation of the PET study. Therefore, OSEM + TOF is the optimum model for tracers with diffuse cortical uptake (i.e. FDG and FMZ), while OSEM + PSF + TOF should be used for tracers with focal uptake, such as MET or FDOPA. Optimum reconstruction parameters have been proposed for each reconstruction algorithm to take the best advantage of PET/CT in the clinical setting. Acknowledgements This work was partially funded by the Government of Spain, Institute of Health Carlos III, Ministry of Science and Innovation (project ADE 10/00028) and by Siemens Healthcare. References [1] Conti M. State of the art and challenges of time-of-flight PET. Phys Med 2009;25:1–11. [2] Panin VY, Kehren F, Michel C, Casey M. Fully 3-D PET reconstruction with system matrix derived from point source measurements. IEEE Trans Med Imaging 2006;25:907–21. [3] Budinger TF. Time-of-flight positron emission tomography: status relative to conventional PET. J Nucl Med 1983;24:73–8. [4] Tong S, Alessio A, Kinahan P. Noise and signal properties in PSF-based fully 3D PET image reconstruction: an experimental evaluation. Phys Med Biol 2010;55:1453–73. [5] Tong S, Alessio AM, Thielemans K, Stearns C, Ross S, Kinahan PE. Properties and mitigation of edge artifacts in PSF-based PET reconstruction. IEEE Trans Nucl Sci 2011;58:2264–75. [6] Surti S, Karp JS, Popescu LM, Daube-Witherspoon ME, Werner M. Investigation of time-of-flight benefit for fully 3-DPET. IEEE Trans Med Imaging 2006;25:529– 38. [7] Lois C, Jakoby BW, Long MJ, Hubner KF, Barker DW, Casey ME, et al. An assessment of the impact of incorporating time-of-flight information into clinical PET/CT imaging. J Nucl Med 2010;51:237–45. [8] Prieto E, Domínguez-Prado I, García-Velloso MJ, Peñuelas I, Richter JA, Martí-Climent JM. Impact of time-of-flight and point-spread-function in SUV quantification for oncological PET. Clin Nucl Med 2013;38:103–9. [9] Akamatsu G, Ishikawa K, Mitsumoto K, Taniguchi T, Ohya N, Baba S, et al. Improvement in PET/CT image quality with a combination of point-spread function and time-of-flight in relation to reconstruction parameters. J Nucl Med 2012;53:1716–22. [10] Nagaki A, Onoguchi M, Matsutomo N. Clinical validation of high-resolution image reconstruction algorithms in brain 18F-FDG-PET: effect of incorporating
Please cite this article in press as: Elena Prieto, et al., Brain PET imaging optimization with time of flight and point spread function modelling, Physica Medica (2015), doi: 10.1016/ j.ejmp.2015.07.001
ARTICLE IN PRESS E. Prieto et al./Physica Medica ■■ (2015) ■■–■■
8
[11]
[12]
[13]
[14]
[15] [16]
[17]
Gaussian filter, point spread function, and time-of-flight. Nucl Med Commun 2014;35:1224–32. Teune LK, Renken RJ, de Jong BM, Willemsen AT, van Osch MJ, Roerdink JB, et al. Parkinson’s disease-related perfusion and glucose metabolic brain patterns identified with PCASL-MRI and FDG-PET imaging. Neuroimage Clin 2014;5:240–4. Walter F, Cloughesy T, Walter MA, Lai A, Nghiemphu P, Wagle N, et al. Impact of 3,4-dihydroxy-6-18F-fluoro-L-phenylalanine PET/CT on managing patients with brain tumors: the referring physician’s perspective. J Nucl Med 2012;53:393–8. Lapa C, Linsenmann T, Monoranu CM, Samnick S, Buck AK, Bluemel C, et al. Comparison of the amino acid tracers 18F-FET and 18F-DOPA in high-grade glioma patients. J Nucl Med 2014;55:1611–16. Martí-Climent JM, Prieto E, Domínguez-Prado I, García-Velloso MJ, Rodríguez-Fraile M, Arbizu J, et al. Contribution of time of flight and point spread function modeling to the performance characteristics of the PET/CT Biograph mCT scanner. Rev Esp Med Nucl Imagen Mol 2013;32:13–21. Hoffman E, Cutler P, Digby W, Mazziotta J. 3-D phantom to simulate cerebral blood flow and metabolic images for PET. IEEE Trans Nucl Sci 1990;37:616–20. Baghaei H, Mawlawi O, Wang Y, Li H, Ramirez R, Kim S, et al. A comparison of five whole-body PET scanners by scanning Hoffman brain phantom. IEEE Nucl Sci Symp Conf Rec 2006;4:1973–6. Freedenberg MI, Badawi RD, Tarantal AF, Cherry SR. Performance and limitations of positron emission tomography (PET) scanners for imaging very low activity sources. Phys Med 2014;30:104–10.
[18] Kadrmas DJ, Casey ME, Conti M, Jakoby BW, Lois C, Townsend DW. Impact of time-of-flight on PET tumor detection. J Nucl Med 2009;50:1315–23. [19] Martí-Climent JM, Prieto E, Elosúa C, Rodríguez-Fraile M, Domínguez-Prado I, Vigil C, et al. PET optimization for improved assessment and accurate quantification of 90Y-microsphere biodistribution after radioembolization. Med Phys 2014;41:092503. [20] Conti M. Effect of randoms on signal-to-noise ratio in TOF PET. IEEE Trans Nucl Sci 2006;53:1188–93. [21] Snyder DL, Miller MI, Thomas LJ, Politte DG. Noise and edge artifacts in maximum-likelihood reconstructions for emission tomography. IEEE Trans Med Imaging 1987;6:228–38. [22] Joshi A, Koeppe RA, Fessler JA. Reducing between scanner differences in multi-center PET studies. Neuroimage 2009;46:154–9. [23] Bettinardi V, Presotto L, Rapisarda E, Picchio M, Gianolli L, Gilardi M. Physical performance of the new hybrid PET/CT Discovery-690. Med Phys 2011;38:5394–411. [24] van Velden FH, Kloet RW, van Berckel BN, Lammertsma AA, Boellaard R. Accuracy of 3-dimensional reconstruction algorithms for the high-resolution research tomograph. J Nucl Med 2009;50:72–80. [25] Ikari Y, Nishio T, Akamatsu G, Miya Y, Senda M. PET camera qualification criteria for brain amyloid imaging: gray/white contrast of Hoffman phantom images. J Nucl Med 2014;55(Suppl. 1):2059.
Please cite this article in press as: Elena Prieto, et al., Brain PET imaging optimization with time of flight and point spread function modelling, Physica Medica (2015), doi: 10.1016/ j.ejmp.2015.07.001