- Email: [email protected]
CT acquisition protocol
European Journal of Radiology 81 (2012) 3363–3370
Contents lists available at SciVerse ScienceDirect
European Journal of Radiology journal homepage: www.elsevier.com/locate/ejrad
Phantom study of the impact of reconstruction parameters on the detection of mini- and micro-volume lesions with a low-dose PET/CT acquisition protocol Alice Ferretti a,b,∗ , Elena Bellan a,1 , Marcello Gava a,1 , Sotirios Chondrogiannis b,2 , Arianna Massaro b,2 , Otello Nibale a,1 , Domenico Rubello b,2 a b
Department of Medical Physics, Santa Maria della Misericordia Hospital, Via Tre Martiri 140, 45100 Rovigo, Italy Department of Nuclear Medicine & PET/CT Centre, Santa Maria della Misericordia Hospital, Via Tre Martiri 140, 45100 Rovigo, Italy
a r t i c l e
i n f o
Article history: Received 23 January 2012 Received in revised form 16 February 2012 Accepted 2 May 2012 Keywords: PET/CT image quality Phantom measurements Partial volume effect Recovery coefficient
a b s t r a c t Purpose: Every PET scanner suffers of the partial volume effect (PVE), that is a loss of contrast in small lesions causing a worsening in standardized uptake value (SUV) accuracy, that is critical if quantitative PET/CT imaging is used for diagnosis and therapy. Methods: In order to quantify PVE and optimize our clinical protocols to minimize this effect in a last generation PET/CT scanner, we utilized a cylindrical phantom equipped with ten mini- and micro-volume hollow spheres. The lesion detectability and the SUV accuracy were evaluated at a fixed spheres to background intrinsic contrast (activity concentration ratio 8:1) but in different scan conditions: (a) acquisition modality (3D vs. 2D), (b) number of subset per iteration, (c) type of post-reconstruction filter and (d) activity concentration (i.e. total counts). Also the effect of different absorber thickness was evaluated. Results: Small lesion detectability resulted better in images acquired in 3D mode rather than 2D, mainly because of the lower noise produced by the fully-3D algorithm. The number of reconstruction iterations and the post-processing filter used affected both the contrast underestimation and the spatial resolution. Decreasing the 18 F activity injected according to the low-dose protocol, the small lesions could be distinguished from the background down to a diameter of 6.2 mm and the SUV accuracy did not deteriorate. Adding absorber thickness around the phantom, the image noise slightly increased while SUV accuracy did not change. Conclusions: The hybrid PET/CT scanner we evaluated showed good performances, mainly in 3D acquisition modality. The phantom measurements showed that the most appropriate reconstruction protocol derived from a compromise between the contrast accuracy and the noise variance in PET images. The lowdose protocol clinically used demonstrated no loss in SUV accuracy and an adequate lesion detectability for lesions down to 6.2 mm in diameter. © 2012 Elsevier Ireland Ltd. All rights reserved.
1. Introduction The positron emission tomography (PET) is increasingly used for the diagnosis and management of several diseases, with
∗ Corresponding author at: Nuclear Medicine Department, Santa Maria della Misericordia Hospital, Via Tre Martiri 140, 45100 Rovigo, Italy. Tel.: +39 0425 39 4430; fax: +39 0425 39 4434. E-mail addresses: [email protected] (A. Ferretti), [email protected] (E. Bellan), [email protected] (M. Gava), [email protected] (S. Chondrogiannis), [email protected] (A. Massaro), [email protected] (O. Nibale), [email protected] (D. Rubello). 1 Tel.: +39 0425 39 3386; fax: +39 0425 39 3386. 2 Tel.: +39 0425 39 4430; fax: +39 0425 39 4434. 0720-048X/$ – see front matter © 2012 Elsevier Ireland Ltd. All rights reserved. http://dx.doi.org/10.1016/j.ejrad.2012.05.001
fundamental applications in oncology and radiotherapy (RT) [1,2]. Usually the PET equipment is associated to a computed tomography (CT) scanner, to allow an accurate correction of the photon attenuation in patient and an easier anatomical localization of the areas of abnormal uptake of the radiopharmaceuticals. The PET images interpretation in fact is based on the knowledge of the radiopharmaceutical distribution in organs and tissues of patient, and physicians often use quantitative information extracted from images, such as the SUV (standardized uptake value), in order to perform a diagnosis, to compare individual cases with the literature and to extract a biological target volume (BTV) for RT planning. For these reasons, particular importance should be given to image quality and quantitative accuracy of PET/CT scans [3]. Because of the limited spatial resolution of the PET scanner, the image blurring causes an expanded but less intense image of small lesions, causing an underestimation of SUV value. This effect, so-called partial volume effect (PVE), increases with decreasing of lesion size. SUV value
3364
A. Ferretti et al. / European Journal of Radiology 81 (2012) 3363–3370
is less than 90% of the true value in lesions smaller than three times the PET resolution [4], calculated as the full width at half maximum (FWHM) of the emission image of a capillary source, with internal diameter of 1 mm. PVE depends on the lesion size, the intrinsic contrast between lesion and background, and the aforementioned spatial resolution [5], which derived from detectors characteristics, reconstruction methods and smoothing filters used. Also the image sampling (i.e. voxelization) produces PVE. Several works were published about the PVE, in particular based on the evaluation of the recovery coefficients (RC), that is the percentage underestimation of the uptake measured in hollow spheres inserted in a cylindrical or elliptic phantom [6–9]. The impact of image noise, spatial resolution and region of interest (ROI) definition on PVE were already evaluated by simulations [8,10]. These data can be used to correct SUV values in clinical images, in order to increase PET/CT quantitative accuracy [11–13]. With particular regards to the hybrid imaging system object of our study (GE Discovery STE PET/CT) its performances according to NEMA NU2-2001 protocol have been already investigated by some authors [6,7,14,15]: the standard PET Body Phantom employed in the above cited investigations is equipped with a set of six spheres of internal diameters (ID) ranging between 10 and 37 mm. Some authors also compare Discovery STE PET/CT image quality in 2D and 3D mode for smaller hot spheres [8,9]. In this work we described measurements of lesion detectability and hot contrast underestimation performed on mini- and microvolume hollow spheres (down to an internal volume of 31 l and corresponding to an ID of 3.9 mm), positioned in a cylindrical phantom. The dependence of lesion detectability on the acquisition mode, total counts collected and reconstruction methods employed was also investigated. The thickness of tissue to be traversed by 511 keV photons affects the PET image quality, and reasonably also the PVE. Also this aspect was investigated with specific measurements.
2. Materials and methods 2.1. PET/CT scanner The measurements were performed with a Discovery STE (General Electric, Milwaukee, WI, USA) PET/CT hybrid scanner installed in our institution. The PET scanner has 24 rings of detectors, subdivided in 8 × 6 blocks of BGO crystals. The bore has a diameter of 70 cm, with an axial field of view (FOV) extension of 15.7 cm, acquiring 47 slices of 3.3 mm thickness. Acquisitions can be obtained both in 2D and 3D mode. Usually only the 3D modality is performed in the clinical setting of our center. The system includes the 16-slices computed tomography scanner Lightspeed16 (General Electric).
2.2. Phantom We focused on measuring the capability to detect very small lesions in PET images, simulated by means of two sets of hollow spheres (Data Spectrum Corporation, Hillsborough, USA) mounted on a standard Jaszczak cylindrical phantom (Data Spectrum Corporation) with 22 cm internal diameter and 19 cm height. The first set is composed of six standard spheres with internal volumes of 16.0, 8.0, 4.0, 2.0, 1.0 and 0.5 ml, respectively, while the micro-volume set includes four spheres with internal volume of 250, 125, 63 and 31 l, respectively. The IDs range between 3.9 and 31.3 mm, with two spheres smaller than PET spatial resolution. The activity concentrations in the hollow spheres and surrounding background were chosen in order to have a lesion to background ratio of 8:1. In order to evaluate the effect of patient
dimensions on PVE, the cylindrical phantom was also inserted in a tank (30 cm × 28 cm × 35 cm) filled with 22 l of water. 2.3. Acquisition protocols The quality of PET images is strongly affected by the acquisition mode (2D or 3D mode), the acquisition duration, the activity concentration (intrinsic contrast), the type of the reconstruction algorithm (actually iterative methods are the standard choice), the number of iterations and the post-reconstruction filter applied. The measurements presented in this work were performed varying some of these parameters. The cylindrical phantom and the spheres were filled with a solution of water and 18 F-FDG. For each measurements session, the background was filled with an activity concentration of 8.3 kBq/ml calibrated at the beginning of the first acquisition. PET scans both in 2D and 3D mode were acquired at different times, with a background activity concentration respectively of: 8.3, 3.7 and 1.6 kBq/ml for PET acquisitions in 2D mode and 7.8, 3.5 and 1.5 kBq/ml for PET acquisitions in 3D mode. In this way we simulated patients injected 1 h before scan start with 11.4–2.2 MBq/kg b.w. The 18 F-FDG concentration in spheres was 66.8 kBq/ml, calibrated at the beginning of the first acquisition. Each scan had the standard duration of 3.5 min in 2D mode and 3 min in 3D mode, as clinically used for whole-body acquisitions at our institute. Finally, in order to measure the spatial resolution of the PET system, we performed a 3 min long PET emissive acquisition in 3D mode of a fine line source positioned at 1 cm distance from the PET isocenter. The capillary tube (internal diameter 1 mm) was filled with 25 MBq of 18 F-FDG, calibrated at the start of the acquisition. 2.4. Reconstruction methods The emissive images were reconstructed with the reconstruction algorithm implemented in the PET/CT scanner for the clinical use: the OSEM (ordered subset expectation maximization) algorithm for images acquired in 2D mode, while GE VUE-Point iterative algorithm for 3D mode scans. The VUE-Point algorithm, a fully-3D OSEM method, performs all the corrections to PET raw data during the iterative loops, and includes a new volumetric scatter correction and an integrated model of the detector geometry which allows an improvement in PET resolution and a reduction in PVE. Initially we performed 3D and 2D reconstructions with 20 subsets and 2 iterations. The post-reconstruction filter for both modalities was set to 5 mm FWHM. The visualization matrix was 256 × 256 pixels (pixel side of 2.7 mm). As regards 2D PET acquisitions, the final image noise resulted higher than 3D, mainly because of the different reconstruction method used. In order to reduce the 2D image noise, we should reduce the number of subsets in the OSEM reconstruction, passing from 20 to 14 subsets, but hazarding to loose the convergence of the algorithm, causing a biased contrast [16]. Thus we chose to compare the two acquisition modalities (2D vs. 3D) using the same number of iterations and subsets. Then, in order to evaluate the effect on the small lesion detectability, we reconstructed the 3D PET raw data sets also: (i) with different numbers of subsets (14 and 28), (ii) with different post-reconstruction filters (3 mm and 7 mm FWHM), (iii) with different reconstruction matrix. The line source image acquired to compute the PET spatial resolution in air, was reconstructed with the same procedure above presented. 2.5. Data analysis The image analysis was performed in a GE workstation provided with the software Advantage Window version 4.4 (ADW4.4), using a dedicated protocol to review the fused PET/CT images, which allowed to calculate quantitative statistical information in terms of
A. Ferretti et al. / European Journal of Radiology 81 (2012) 3363–3370
local uptake L (kBq/ml) or of SUV (g/ml), according to the equation:
3365
60%
(a)
3D - VUE Point - 2 iter. x 20 subs.
L (kBq/ml) A (kBq)/b.w. (g)
where L is the local uptake of 18 F in the attenuation-corrected PET images, b.w. is the body weight of the patient (or of the phantom), and A is the 18 F activity at the start of PET acquisition. The conversion from counts to local uptake L (kBq/ml) is performed by the GE image elaboration software using the PET calibration data (i.e. well-counter calibration). In particular the parameters used in subsequent calculations were: Lmean,i and LSD,i , respectively the measured mean uptake and the standard deviation in each lesion, both calculated in ROIs of diameter close to PET resolution (area 30.8 mm2 ), and Mbkg and MSD , respectively the measured mean uptake and the mean standard deviation of the background, calculated using six circular ROIs with an area of 870 mm2 positioned in the uniform areas of the phantom axial images. These values were used to evaluate the noise, the partial volume effect and the small lesion detectability, computing:
50%
background CV
SUV =
2D - OSEM - 2 iter. x 20 subs.
40%
30%
20%
10%
0% 5
0
10
15
20
25
30
35
40
45
19.7
24.8
31.3
slice position 50
(b) 2D - OSEM 2 iter. x 20 subs. 3D - VUE-Point 2 iter. x 20 subs.
30
LOBV
a) the coefficient of variation, defined as CV = Mbkg /MSD , to quantify image noise and its variability intra-slice; b) the lesion over background variation, defined according to the Rose Model of statistical detection [17] as LOBV = (Lmean − Mbkg )/(k×MSD ), where k is a constant value. LOBV value greater than 1 corresponds to a lesion distinguishable over the background, for a human observer, with a fixed confidence level. In PET imaging [6,7], a constant k value equal to 3 is often used assuming a Gaussian noise distribution, obtaining a confidence level of 99.73%, as the product k×MSD results equal to three standard deviations; c) the recovery coefficients defined as RC = (Lmean /Mbkg )/(Clesion /Cbkg ) where Clesion and Cbkg are the real activity concentrations present in the lesion and in the background [18]. RC represents the percentage of contrast detected.
40
20
10
0
3.9
4.9
6.2
7.8
9.8
12.4
15.6
internal diameter (mm) 120%
(c)
3. Results 100%
3.1. Small lesion detectability for 2D and 3D acquisition mode (standard reconstruction)
RC
80%
As shown in Fig. 1a, in the 30 central slices, the mean coefficient of variation in 2D mode resulted twice higher than in 3D mode, while in the external slices this discrepancy decreased. The lower noise produced during 3D reconstruction (mean CV 6% in the central part of the axial FOV for an acquisition of 3 min with 7.8 kBq/ml) can be attributed to the high-performance of the fully3D reconstruction algorithm implemented by GE. The typical 3D sensitivity curve as a function of slice position (with triangular shape) explains the worsening in the external planes. For our subsequent measurements, the phantom was centered in the scanner so that the hot spheres were imagined with central rings of detectors, corresponding to slices ranging between 25 and 35. In Fig. 1b LOBV values versus spheres ID in the two acquisition modalities are shown, at similar concentration of activity in the phantom background (3.5 kBq/ml in 3D and 3.7 kBq/ml in 2D acquisition). In 3D mode, lesion detectability resulted clearly higher than 2D mode, because of the lower noise present in 3D images, as shown in Fig. 1a. The RC values calculated in the same images are shown in Fig. 1c.
60%
40%
2D - OSEM 2 iter. x 20 subs.
20%
3D - VUE-Point 2 iter. x 20 subs.
0% 0
5
10
15
20
25
30
35
internal diameter (mm) Fig. 1. (a) The background CV versus the slice number, (b) the LOBV values and (c) the RC values versus hot spheres ID (intrinsic contrast 8:1), comparing 3D and 2D acquisition mode, respectively at 7.8 and 8.3 kBq/ml activity concentration in the background. For 2D images the number of subsets was set to 14 in order to reduce the CV near the values measured in 3D reconstructed images.
3366
A. Ferretti et al. / European Journal of Radiology 81 (2012) 3363–3370
3.2. Effect of the reconstruction parameters
3.3. Effect of phantom dimension
(a) 3D - VUE-Point 2 iter. x 14 subs.
50%
3D - VUE-Point 2 iter. x 20 subs.
background CV
3D - VUE-Point 2 iter. x 28 subs.
40%
30%
20%
10%
0% 0
5
10
15
20
25
30
35
40
45
15.6
19.7
24.8
31.3
slice position 60
(b)
3D - VUE-Point 2 iter. x 14 subs. 3D - VUE-Point 2 iter. x 20 subs.
50
3D - VUE-Point 2 iter. x 28 subs.
LOBV
40
30
20
10
0 3.9
4.9
6.2
7.8
9.8
12.4
internal diameter (mm) 120%
(c) 100%
80%
RC
According to the references [19,20], the ordered-subset iterative reconstruction reaches the convergence for large values of the product of subsets by iterations (product greater or equal to 50). It was already shown [16] that a reduction in this product generates images with a low noise level but at the expense of a biased contrast, in particular in low-count regions. In order to reach a good compromise between the contrast accuracy and a low variance, we compared PET images, acquired in 3D mode, reconstructed with the clinically implemented method (20 subsets and 2 iterations) with the same raw data reconstructed with two different number of subsets (14 and 28). In Fig. 2a the mean CV values measured in a circular ROI (8.7 cm2 ) positioned in the central area of the phantom are shown for the three different number of subsets, while the calculated LOBV and RC values for these reconstruction modalities are shown in Fig. 2b and c for the PET/CT images with 7.8 kBq/ml in background. Increasing the product subsets by iterations, images resulted noiser, and so the LOBV values decreased, while the measured contrast (i.e. RC values) improved. In Fig. 3a the mean CV values and the measured spatial resolution in air, for different post-reconstruction filters (3, 5 and 7 mm FWHM) are shown. The increased post-processing filter of 7 mm FWHM caused a smoothing in images (i.e. worsening in spatial resolution), while it reduced the noise variance (i.e. CV values). The impact on PVE is shown in Fig. 3b and c. As regards LOBV values (Fig. 3b), stronger filter caused an improvement in lesion detectability, particularly evident passing from 3 to 5 mm FWHM filter. Conversely, the contrast underestimation (Fig. 3c) had an inverse trend, worsening when the filter grew. In order to test the combined effect of the number of iterations and the post-reconstruction smoothing, we compared the subsequent cases: (i) 14 subsets × 2 iterations and 3 mm FWHM smoothing, (ii) 20 subsets × 2 iterations and 5 mm smoothing, (iii) 28 subsets × 2 iterations and 7 mm smoothing, as reported in Fig. 4. Finally, the impact of the reconstruction matrix on PVE is shown in Fig. 5a–c. As CV values did not change significantly from 128 × 128 to 256 × 256 reconstruction matrix (Fig. 5a), the lesion detectability was quite constant for the bigger spheres (Fig. 5b). A relevant increase in measured contrast was recorded with the finest matrix for spheres ID under 2 cm, together with an increased lesion detectability (Fig. 5c). Also the Nyquist sampling theorem substantiated this findings. According to it, the optimal pixel size is equal to half the spatial resolution. A greater pixel size produces undersampling and so aliasing, while smaller pixel size produces useless oversampling. Thus, the 256 × 256 matrix, with pixel size of 2.7 mm coupled to our post-reconstruction filter at 5 mm FWHM, resulted in the optimal choice.
60%
60%
40% 3D - VUE-Point 2 iter. x 14 subs.
A further aspect investigated regards the patient dimensions. In order to evaluate PVE dependence on absorber thickness, we performed further acquisitions inserting our cylindrical phantom in a water tank, increasing the water thickness of 10 cm laterally and of 7 cm vertically. The additional scattering and attenuation of photons produced a slight increase in noise (see Fig. 6a), and consequently lower values of LOBV (Fig. 6b). Conversely, the RC values did not changed significantly, as shown in Fig. 6c. 3.4. The low-dose acquisition protocol In our institution a low dose protocol was adopted for 18 F-FDG PET/CT exams, injecting to each patient 2.2 MBq/kg, in order to
3D - VUE-Point 2 iter. x 20 subs.
20%
3D - VUE-Point 2 iter. x 28 subs.
0% 0
10
20
30
40
internal diameter (mm) Fig. 2. (a) The mean CV, (b) the LOBV values and (c) the RC values versus spheres internal diameters (intrinsic contrast 8:1), comparing 3D data acquired at an activity concentration of 8.3 kBq/ml and reconstructed by VUE-Point algorithm with 2 iterations but different number of subsets (14, 20 and 28).
A. Ferretti et al. / European Journal of Radiology 81 (2012) 3363–3370
12%
12
3367
120%
(a) mean CV (left axis) FWHM (right axis)
100%
10 10%
8%
RC
FWHM
mean CV
80% 8
60%
6
40% 6% 4 3D - VUE-Point 2x14 filter 3mm 3D - VUE-Point 2x20 filter 5mm 3D - VUE VUE-Poin Pointt 2x2 2x28 8 fil filter ter 7m 7mm m
20% 4%
2 3mm
5mm
7mm
0%
post-reconstruction filter
0
(b) 50
3D - VUE-Point filter 3mm 3D - VUE-Point filter 5mm
20
30
40
Fig. 4. The RC values versus spheres ID (intrinsic contrast 8:1) comparing 3D data acquired at an activity concentration of 8.3 kBq/ml and reconstructed with different combinations of iterations and smoothing.
3D - VUE-Point filter 7mm
40
LOBV
10
internal diameter (mm)
60
decrease patient internal exposure. CT and PET images are acquired 1 h later, with an activity reduced to 1.5 MBq/kg (equivalent to 1.5 kBq/ml in our cylindrical phantom). To quantify the small lesion detectability and the SUV accuracy for this protocol, we acquired PET scans of the same Jaszczak phantom exploiting the 18 F isotope decay, decreasing progressively the concentration of radioactivity and consequently the count rate. Spheres to background intrinsic contrast was kept always equal to 8:1. The LOBV values versus spheres ID, obtained decreasing the activity, are reported in Fig. 7a for the six smaller spheres acquired in 3D mode at different activity concentrations of the background. Clearly the LOBV values increased with the activity (that is with counts rate), but the lesions can be distinguish from the background (LOBV greater than 1) starting from a diameter of 6.2 mm. In addition, Fig. 7b confirms that adopting this low dose protocol, the SUV accuracy does not deteriorate. In fact RC versus spheres diameters followed the same curve at different activities.
30
20
10
0 3.9
4.9
6.2
7.8
9.8
12.4
15.6
19.7
24.8
31.3
internal diameter (mm) 120%
(c) 100%
4. Discussion
RC
80%
60%
40% 3D - VUE-Point filter 3mm 3D - VUE-Point filter 5mm
20%
3D - VUE-Point filter 7mm
0% 0
10
20
30
40
internal diameter (mm) Fig. 3. (a) The mean CV and the spatial resolution measure in air, (b) the LOBV values and (c) the RC values versus spheres ID (intrinsic contrast 8:1), comparing 3D data acquired at an activity concentration of 7.8 kBq/ml and reconstructed applying three different smoothing filter (3, 5 and 7 mm FWHM).
Many studies were performed to evaluate the PVE in PET/CT scanners, both for 2D and 3D acquisition modalities. Usually those measurements were performed with an elliptical body phantom equipped with hollow spheres with ID ranging from 10 to 37 mm [4–7]. Some authors also measured the SUV underestimation in PET images by means of smaller spheres [8,9]. Also the impact of intrinsic contrast of lesions over background, image noise and spatial resolution on SUV accuracy was calculated by computer simulations [9,10]. Aim of the present study was to optimize the acquisition and reconstruction protocols clinically used with patients in our center, in order to maximize the SUV accuracy and the detectability of lesions with dimensions near the PET spatial resolution. A good compromise between PET quantitative accuracy and noise variance (quantities in competition during the iterative reconstruction process) was reached acquiring PET/CT images in 3D mode and reconstructing them with VUE-Point algorithm (a fully-3D OSEMtype iterative method). We investigated the impact on PVE of the reconstruction process, in particular as regards: (a) the overall iterations used (result of the product between subsets number
3368
A. Ferretti et al. / European Journal of Radiology 81 (2012) 3363–3370
60%
60%
(a)
(a)
phantom only
128x128 3D - VUE-Point 256x256 3D - VUE-Point
50%
background CV
40%
background CV
phantom with additional thickness
50%
30%
40%
30%
20%
20%
10%
10%
0%
0% 0
5
10
15
20
25
30
35
40
5
0
45
10
50
45 3D - VUE-Point 256x256
40
25
30
35
40
45
15.6
19.7
24.8
31.3
phantom only
3D - VUE-Point 128x128
40
20
(b)
(b) 45
15
slice position
slice position
phantom with additional thickness
35
35
30
LOBV
LOBV
30 25
25 20
20 15
15 10
10
5
5
0
0 3.9
4.9
6.2
7.8
9.8
12.4
15.6
19.7
24.8
3.9
31.3
4.9
6.2
9.8
12.4
120%
120%
(c)
(c) 100%
100%
80%
RC
80%
RC
7.8
internal diameter (mm)
internal diameter (mm)
60%
60%
40%
40%
phantom only
20%
3D - VUE-Point - 256x256
20%
phantom with additional thickness
3D - VUE-Point - 128x128
0% 0%
0 0
10
20
30
40
internal diameter (mm) Fig. 5. (a) The background CV versus slice position, (b) the LOBV and (c) RC values versus spheres ID, comparing two reconstruction matrixes (128 × 128 and 256 × 256 pixel), for 3D VUE-Point reconstructions at 7.8 kBq/ml activity concentration in the background and 8:1 hot spheres intrinsic contrast.
10
20
30
40
internal diameter (mm) Fig. 6. (a) The background CV versus slice position, (b) the LOBV and (c) RC values versus spheres ID, with and without an additional absorber (inserting the cylindrical phantom in a water tank 30 cm × 28 cm × 35 cm), acquired in 3D mode at 7.8 kBq/ml activity concentration in the background and 8:1 hot spheres intrinsic contrast.
A. Ferretti et al. / European Journal of Radiology 81 (2012) 3363–3370
20
The limit of lesion detectability was identified about 6–7 mm lesion diameter, also with the low-activity protocol adopted in our institute. This low-dose protocol (2.2 MBq/kg b.w. 18 F-FDG activity injected to patients) assured a low internal exposition to the patient and an unchanged SUV accuracy, as shown by the present measurements. The whole-body PET/CT image quality was evaluated by experienced nuclear medicine physicians in clinical images and it resulted adequate, allowing to keep restrained the exam duration (about 20 min for a whole-body scan) and reducing the internal exposition of patients. Furthermore, we tested how the quantitative accuracy changed with patients thickness, repeating the measurements inserting the phantom in an additional container (30 cm × 28 cm × 35 cm) filled with water. We found no significantly effects on SUV accuracy (i.e. RC values), but the small lesion detectability over the background was reduced. Clinically, these data suggested some further caution in the interpretation of PET/CT images in obese patients due to a slight increase of the noise. An issue that was not treated in this work regards the implementation of the PVE correction, that requires further phantom measurements for different values of intrinsic lesion-to-background contrast. This aspect remains an open issue, to be developed soon.
(a) 3D - 1.5 kBq/ml - 3 min 3D - 3.5 kBq/ml - 3 min 3D - 7.8 kBq/ml - 3 min
LOBV
15
10
5
0 3.9
4.9
6.2
7.8
9.8
3369
12.4
internal diameter (mm) 120%
(b) 100%
5. Conclusions
RC
80%
60%
40% 3D - 1.5 kBq/ml - 3 min 3D - 3.5 kBq/ ml - 3 min
20%
3D - 7.8 kBq/ml - 3 min
0%
0
10
20
30
40
internal diameter (mm) Fig. 7. (a) The LOBV values of the smallest hot spheres and (b) the RC values versus spheres internal diameter, for images acquired at three different values of activity concentration in the background (3D VUE-Point reconstruction).
and iterations number) and (b) the post-processing filter used to reduce the image noise. As regards the first issue, in literature the debate about the minimum number of overall iterations necessary to assure the iteration algorithm convergence is already open [16]. We chose to measure the effect comparing three cases, reconstructing the same PET raw data with 14, 20 and 28 subsets and 2 iterations, respectively corresponding to a total of 28, 40 and 56 iterations in the expectation–maximization method. Decreasing the number of subsets, the noise variance decrease, increasing the lesion detectability (i.e. LOBV values). Conversely, the RC curve versus spheres ID became less steep and so the measured contrast worsen for lesions ranging from 8 to 20 mm. As regards the second issue, the post-processing filter is used to reduce the image noise produced by the iterative algorithm. An additional smoothing, coupled with a high number of iterations, could reduce the noise variance, but producing a biased lesion-to-background contrast. We tested also the possibility to vary both the number of iterations and the smoothing filter, concluding that images obtained with 20 subsets × 2 iterations and 5 mm FWHM postreconstruction filter reached a good compromise between lesion detectability over the background and contrast accuracy.
The PET/CT scanner GE Discovery STE showed good performances, mainly in 3D acquisition modality, as regards both the detection of micro-volume hot spheres and the noise variance. The data shown in this work were collected using a set of very small hollow spheres, even smaller than the resolution of the PET scanner, with a sphere to background intrinsic contrast of 8:1, similar to clinical SUV values found in patients. Present measurements showed that hot spheres greater than 6 mm resulted detectable also simulating the low dose protocol used in clinical cases (2.2 MBq/kg injected 1 h before scan start). The partial volume effect was quantified by means of the RC. These values decreased under 90% for spheres ID smaller or equal to 12.4 mm, and it is about 50% for spheres internal diameter of 10 mm. RC values appeared to be not dependent on the concentration of activity present in the spheres at fixed scan duration (i.e. the total counts acquired), but they are affected by the number of subsets of the iterative reconstruction and the post-processing smoothing filter used. In particular these parameters had an inverse impact on RC and LOBV: measured contrast improved while noise and LOBV values worsened when the number of subsets were increased or the smoothing was reduced. The standard 3D scan reconstructed by VUE-Point algorithm with 20 subsets × 2 iterations and 5 mm FWHM filter appeared to be a good compromise between low noise variance and high SUV accuracy in small lesions. Even if the scatter and random coincidences acquired in 2D mode are less than in 3D mode, the improved VUE-Point iterative algorithm available in our hybrid scanner for fully-3D reconstructions produces PET images with a very low noise level and a good quantitative accuracy, with a final image quality higher than 2D-OSEM PET images produced by the same system.
References [1] Townsend DW. Positron emission tomography/computed tomography. Seminars in Nuclear Medicine 2008;38:152–66. [2] Sattler B, Lee JA, Lonsdale M, Coche E. PET/CT (and CT) instrumentation, image reconstruction and data transfer for radiotherapy planning. Radiotherapy and Oncology 2010;96:288–97. [3] Weber WA. Quantitative analysis of PET studies. Radiotherapy and Oncology 2010;96:308–10. [4] Geworski L, Knoop BO, Levi de Cabreias M, Knapp WH, Munz DL. Recovery correction for quantitation in emission tomography: a feasibility study. European Journal of Nuclear Medicine 2000;27:161–9.
3370
A. Ferretti et al. / European Journal of Radiology 81 (2012) 3363–3370
[5] Soret M, Bacharach SL, Buvat I. Partial-volume effect in PET tumor imaging. Journal of Nuclear Medicine 2007;48:932–45. [6] Brown C, Dempsey MF, Gilen G, Elliott AT. Investigation of 18 F-FDG 3D mode PET image quality versus acquisition time. Nuclear Medicine Communications 2001;31:254–9. [7] Bettinardi V, Mancosu P, Danna M, et al. Two-dimensional vs three-dimensional imaging in whole body oncologic PET/CT: a Discovery STE phantom and patient study. Quarterly Journal of Nuclear Medicine and Molecular Imaging 2007;51:214–23. [8] Raylman RR, Kison PV, Wahl RL. Capabilities of two- and three-dimensional FDG-PET for detecting small lesions and lymph nodes in the upper torso: a dynamic phantom study. European Journal of Nuclear Medicine 1999;26:39–45. [9] Brambilla M, Matheoud R, Secco C, et al. Impact of target-to-background ratio, target size, emission scan duration and activity on physical figures of merit for a 3D LSO-based whole body PET/CT scanner. Medical Physics 2007;34:3854–65. [10] Boellaard R, Krak NC, Hoekstra OS, Lammertsma AA. Effect of noise, image resolution, and ROI definition on the accuracy of standard uptake values: a simulation study. Journal of Nuclear Medicine 2004;45:1519–27. [11] Teo BK, Seo Y, Bacharach SL, et al. Partial-volume correction in PET: validation of an iterative postreconstruction method with phantom and patient data. Journal of Nuclear Medicine 2007;48:802–10. [12] Srinivas SM, Dhurairaj T, Basu S, Bural G, Surti S, Alavi A. A recovery coefficient method for partial volume correction of PET images. Annals of Nuclear Medicine 2009;23:341–8.
[13] Barbee1 DL, Flynn RT, Holden JE, Nickles RJ, Jeraj R. Partial volume correction of PET-imaged tumour heterogeneity using expectation maximization with a spatially varying point spread function. Physics in Medicine & Biology 2010;55:221–36. [14] Macdonald LR, Schmitz RE, Alessio AM, et al. Measured count-rate performance of the Discovery STE PET/CT scanner in 2D, 3D and partial collimation acquisition modes. Physics in Medicine & Biology 2008;53:3723–38. [15] Teras M, Tolvanen T, Johansson JJ, Williams JJ, Knuuti J. Performance of the new generation of whole-body PET/CT scanners: Discovery STE and Discovery VCT. European Journal of Nuclear Medicine and Molecular Imaging 2007;34:1683–92. [16] Soret A, Boellaard R, van der Weerdt A. Authors of the letter and the reply – number of iterations when comparing MLEM/OSEM with FBP. Journal of Nuclear Medicine 2004;45:2125–6. [17] Burgess AE. The Rose model, revised. Journal of the Optical Society of America 1999;16:633–46. [18] International Electrotechnical Commission, IEC 61675-1, Radionuclide imaging devices. Characteristics and test conditions. Part 1: Positron emission tomographs; 1998. [19] Boellaard R, O’Doherty MJ, Weber WA, et al. EANM oncology committee and EANM physics committee, FDG PET and PET/CT: EANM procedure guidelines for tumour PET imaging: version 1.0. European Journal of Nuclear Medicine and Molecular Imaging 2010;37:181–200. [20] Hutton BF, Hudson HM, Beekman FJ. A clinical perspective of accelerated statistical reconstruction. European Journal of Nuclear Medicine 1997;24:797–808.
Contents lists available at SciVerse ScienceDirect
European Journal of Radiology journal homepage: www.elsevier.com/locate/ejrad
Phantom study of the impact of reconstruction parameters on the detection of mini- and micro-volume lesions with a low-dose PET/CT acquisition protocol Alice Ferretti a,b,∗ , Elena Bellan a,1 , Marcello Gava a,1 , Sotirios Chondrogiannis b,2 , Arianna Massaro b,2 , Otello Nibale a,1 , Domenico Rubello b,2 a b
Department of Medical Physics, Santa Maria della Misericordia Hospital, Via Tre Martiri 140, 45100 Rovigo, Italy Department of Nuclear Medicine & PET/CT Centre, Santa Maria della Misericordia Hospital, Via Tre Martiri 140, 45100 Rovigo, Italy
a r t i c l e
i n f o
Article history: Received 23 January 2012 Received in revised form 16 February 2012 Accepted 2 May 2012 Keywords: PET/CT image quality Phantom measurements Partial volume effect Recovery coefficient
a b s t r a c t Purpose: Every PET scanner suffers of the partial volume effect (PVE), that is a loss of contrast in small lesions causing a worsening in standardized uptake value (SUV) accuracy, that is critical if quantitative PET/CT imaging is used for diagnosis and therapy. Methods: In order to quantify PVE and optimize our clinical protocols to minimize this effect in a last generation PET/CT scanner, we utilized a cylindrical phantom equipped with ten mini- and micro-volume hollow spheres. The lesion detectability and the SUV accuracy were evaluated at a fixed spheres to background intrinsic contrast (activity concentration ratio 8:1) but in different scan conditions: (a) acquisition modality (3D vs. 2D), (b) number of subset per iteration, (c) type of post-reconstruction filter and (d) activity concentration (i.e. total counts). Also the effect of different absorber thickness was evaluated. Results: Small lesion detectability resulted better in images acquired in 3D mode rather than 2D, mainly because of the lower noise produced by the fully-3D algorithm. The number of reconstruction iterations and the post-processing filter used affected both the contrast underestimation and the spatial resolution. Decreasing the 18 F activity injected according to the low-dose protocol, the small lesions could be distinguished from the background down to a diameter of 6.2 mm and the SUV accuracy did not deteriorate. Adding absorber thickness around the phantom, the image noise slightly increased while SUV accuracy did not change. Conclusions: The hybrid PET/CT scanner we evaluated showed good performances, mainly in 3D acquisition modality. The phantom measurements showed that the most appropriate reconstruction protocol derived from a compromise between the contrast accuracy and the noise variance in PET images. The lowdose protocol clinically used demonstrated no loss in SUV accuracy and an adequate lesion detectability for lesions down to 6.2 mm in diameter. © 2012 Elsevier Ireland Ltd. All rights reserved.
1. Introduction The positron emission tomography (PET) is increasingly used for the diagnosis and management of several diseases, with
∗ Corresponding author at: Nuclear Medicine Department, Santa Maria della Misericordia Hospital, Via Tre Martiri 140, 45100 Rovigo, Italy. Tel.: +39 0425 39 4430; fax: +39 0425 39 4434. E-mail addresses: [email protected] (A. Ferretti), [email protected] (E. Bellan), [email protected] (M. Gava), [email protected] (S. Chondrogiannis), [email protected] (A. Massaro), [email protected] (O. Nibale), [email protected] (D. Rubello). 1 Tel.: +39 0425 39 3386; fax: +39 0425 39 3386. 2 Tel.: +39 0425 39 4430; fax: +39 0425 39 4434. 0720-048X/$ – see front matter © 2012 Elsevier Ireland Ltd. All rights reserved. http://dx.doi.org/10.1016/j.ejrad.2012.05.001
fundamental applications in oncology and radiotherapy (RT) [1,2]. Usually the PET equipment is associated to a computed tomography (CT) scanner, to allow an accurate correction of the photon attenuation in patient and an easier anatomical localization of the areas of abnormal uptake of the radiopharmaceuticals. The PET images interpretation in fact is based on the knowledge of the radiopharmaceutical distribution in organs and tissues of patient, and physicians often use quantitative information extracted from images, such as the SUV (standardized uptake value), in order to perform a diagnosis, to compare individual cases with the literature and to extract a biological target volume (BTV) for RT planning. For these reasons, particular importance should be given to image quality and quantitative accuracy of PET/CT scans [3]. Because of the limited spatial resolution of the PET scanner, the image blurring causes an expanded but less intense image of small lesions, causing an underestimation of SUV value. This effect, so-called partial volume effect (PVE), increases with decreasing of lesion size. SUV value
3364
A. Ferretti et al. / European Journal of Radiology 81 (2012) 3363–3370
is less than 90% of the true value in lesions smaller than three times the PET resolution [4], calculated as the full width at half maximum (FWHM) of the emission image of a capillary source, with internal diameter of 1 mm. PVE depends on the lesion size, the intrinsic contrast between lesion and background, and the aforementioned spatial resolution [5], which derived from detectors characteristics, reconstruction methods and smoothing filters used. Also the image sampling (i.e. voxelization) produces PVE. Several works were published about the PVE, in particular based on the evaluation of the recovery coefficients (RC), that is the percentage underestimation of the uptake measured in hollow spheres inserted in a cylindrical or elliptic phantom [6–9]. The impact of image noise, spatial resolution and region of interest (ROI) definition on PVE were already evaluated by simulations [8,10]. These data can be used to correct SUV values in clinical images, in order to increase PET/CT quantitative accuracy [11–13]. With particular regards to the hybrid imaging system object of our study (GE Discovery STE PET/CT) its performances according to NEMA NU2-2001 protocol have been already investigated by some authors [6,7,14,15]: the standard PET Body Phantom employed in the above cited investigations is equipped with a set of six spheres of internal diameters (ID) ranging between 10 and 37 mm. Some authors also compare Discovery STE PET/CT image quality in 2D and 3D mode for smaller hot spheres [8,9]. In this work we described measurements of lesion detectability and hot contrast underestimation performed on mini- and microvolume hollow spheres (down to an internal volume of 31 l and corresponding to an ID of 3.9 mm), positioned in a cylindrical phantom. The dependence of lesion detectability on the acquisition mode, total counts collected and reconstruction methods employed was also investigated. The thickness of tissue to be traversed by 511 keV photons affects the PET image quality, and reasonably also the PVE. Also this aspect was investigated with specific measurements.
2. Materials and methods 2.1. PET/CT scanner The measurements were performed with a Discovery STE (General Electric, Milwaukee, WI, USA) PET/CT hybrid scanner installed in our institution. The PET scanner has 24 rings of detectors, subdivided in 8 × 6 blocks of BGO crystals. The bore has a diameter of 70 cm, with an axial field of view (FOV) extension of 15.7 cm, acquiring 47 slices of 3.3 mm thickness. Acquisitions can be obtained both in 2D and 3D mode. Usually only the 3D modality is performed in the clinical setting of our center. The system includes the 16-slices computed tomography scanner Lightspeed16 (General Electric).
2.2. Phantom We focused on measuring the capability to detect very small lesions in PET images, simulated by means of two sets of hollow spheres (Data Spectrum Corporation, Hillsborough, USA) mounted on a standard Jaszczak cylindrical phantom (Data Spectrum Corporation) with 22 cm internal diameter and 19 cm height. The first set is composed of six standard spheres with internal volumes of 16.0, 8.0, 4.0, 2.0, 1.0 and 0.5 ml, respectively, while the micro-volume set includes four spheres with internal volume of 250, 125, 63 and 31 l, respectively. The IDs range between 3.9 and 31.3 mm, with two spheres smaller than PET spatial resolution. The activity concentrations in the hollow spheres and surrounding background were chosen in order to have a lesion to background ratio of 8:1. In order to evaluate the effect of patient
dimensions on PVE, the cylindrical phantom was also inserted in a tank (30 cm × 28 cm × 35 cm) filled with 22 l of water. 2.3. Acquisition protocols The quality of PET images is strongly affected by the acquisition mode (2D or 3D mode), the acquisition duration, the activity concentration (intrinsic contrast), the type of the reconstruction algorithm (actually iterative methods are the standard choice), the number of iterations and the post-reconstruction filter applied. The measurements presented in this work were performed varying some of these parameters. The cylindrical phantom and the spheres were filled with a solution of water and 18 F-FDG. For each measurements session, the background was filled with an activity concentration of 8.3 kBq/ml calibrated at the beginning of the first acquisition. PET scans both in 2D and 3D mode were acquired at different times, with a background activity concentration respectively of: 8.3, 3.7 and 1.6 kBq/ml for PET acquisitions in 2D mode and 7.8, 3.5 and 1.5 kBq/ml for PET acquisitions in 3D mode. In this way we simulated patients injected 1 h before scan start with 11.4–2.2 MBq/kg b.w. The 18 F-FDG concentration in spheres was 66.8 kBq/ml, calibrated at the beginning of the first acquisition. Each scan had the standard duration of 3.5 min in 2D mode and 3 min in 3D mode, as clinically used for whole-body acquisitions at our institute. Finally, in order to measure the spatial resolution of the PET system, we performed a 3 min long PET emissive acquisition in 3D mode of a fine line source positioned at 1 cm distance from the PET isocenter. The capillary tube (internal diameter 1 mm) was filled with 25 MBq of 18 F-FDG, calibrated at the start of the acquisition. 2.4. Reconstruction methods The emissive images were reconstructed with the reconstruction algorithm implemented in the PET/CT scanner for the clinical use: the OSEM (ordered subset expectation maximization) algorithm for images acquired in 2D mode, while GE VUE-Point iterative algorithm for 3D mode scans. The VUE-Point algorithm, a fully-3D OSEM method, performs all the corrections to PET raw data during the iterative loops, and includes a new volumetric scatter correction and an integrated model of the detector geometry which allows an improvement in PET resolution and a reduction in PVE. Initially we performed 3D and 2D reconstructions with 20 subsets and 2 iterations. The post-reconstruction filter for both modalities was set to 5 mm FWHM. The visualization matrix was 256 × 256 pixels (pixel side of 2.7 mm). As regards 2D PET acquisitions, the final image noise resulted higher than 3D, mainly because of the different reconstruction method used. In order to reduce the 2D image noise, we should reduce the number of subsets in the OSEM reconstruction, passing from 20 to 14 subsets, but hazarding to loose the convergence of the algorithm, causing a biased contrast [16]. Thus we chose to compare the two acquisition modalities (2D vs. 3D) using the same number of iterations and subsets. Then, in order to evaluate the effect on the small lesion detectability, we reconstructed the 3D PET raw data sets also: (i) with different numbers of subsets (14 and 28), (ii) with different post-reconstruction filters (3 mm and 7 mm FWHM), (iii) with different reconstruction matrix. The line source image acquired to compute the PET spatial resolution in air, was reconstructed with the same procedure above presented. 2.5. Data analysis The image analysis was performed in a GE workstation provided with the software Advantage Window version 4.4 (ADW4.4), using a dedicated protocol to review the fused PET/CT images, which allowed to calculate quantitative statistical information in terms of
A. Ferretti et al. / European Journal of Radiology 81 (2012) 3363–3370
local uptake L (kBq/ml) or of SUV (g/ml), according to the equation:
3365
60%
(a)
3D - VUE Point - 2 iter. x 20 subs.
L (kBq/ml) A (kBq)/b.w. (g)
where L is the local uptake of 18 F in the attenuation-corrected PET images, b.w. is the body weight of the patient (or of the phantom), and A is the 18 F activity at the start of PET acquisition. The conversion from counts to local uptake L (kBq/ml) is performed by the GE image elaboration software using the PET calibration data (i.e. well-counter calibration). In particular the parameters used in subsequent calculations were: Lmean,i and LSD,i , respectively the measured mean uptake and the standard deviation in each lesion, both calculated in ROIs of diameter close to PET resolution (area 30.8 mm2 ), and Mbkg and MSD , respectively the measured mean uptake and the mean standard deviation of the background, calculated using six circular ROIs with an area of 870 mm2 positioned in the uniform areas of the phantom axial images. These values were used to evaluate the noise, the partial volume effect and the small lesion detectability, computing:
50%
background CV
SUV =
2D - OSEM - 2 iter. x 20 subs.
40%
30%
20%
10%
0% 5
0
10
15
20
25
30
35
40
45
19.7
24.8
31.3
slice position 50
(b) 2D - OSEM 2 iter. x 20 subs. 3D - VUE-Point 2 iter. x 20 subs.
30
LOBV
a) the coefficient of variation, defined as CV = Mbkg /MSD , to quantify image noise and its variability intra-slice; b) the lesion over background variation, defined according to the Rose Model of statistical detection [17] as LOBV = (Lmean − Mbkg )/(k×MSD ), where k is a constant value. LOBV value greater than 1 corresponds to a lesion distinguishable over the background, for a human observer, with a fixed confidence level. In PET imaging [6,7], a constant k value equal to 3 is often used assuming a Gaussian noise distribution, obtaining a confidence level of 99.73%, as the product k×MSD results equal to three standard deviations; c) the recovery coefficients defined as RC = (Lmean /Mbkg )/(Clesion /Cbkg ) where Clesion and Cbkg are the real activity concentrations present in the lesion and in the background [18]. RC represents the percentage of contrast detected.
40
20
10
0
3.9
4.9
6.2
7.8
9.8
12.4
15.6
internal diameter (mm) 120%
(c)
3. Results 100%
3.1. Small lesion detectability for 2D and 3D acquisition mode (standard reconstruction)
RC
80%
As shown in Fig. 1a, in the 30 central slices, the mean coefficient of variation in 2D mode resulted twice higher than in 3D mode, while in the external slices this discrepancy decreased. The lower noise produced during 3D reconstruction (mean CV 6% in the central part of the axial FOV for an acquisition of 3 min with 7.8 kBq/ml) can be attributed to the high-performance of the fully3D reconstruction algorithm implemented by GE. The typical 3D sensitivity curve as a function of slice position (with triangular shape) explains the worsening in the external planes. For our subsequent measurements, the phantom was centered in the scanner so that the hot spheres were imagined with central rings of detectors, corresponding to slices ranging between 25 and 35. In Fig. 1b LOBV values versus spheres ID in the two acquisition modalities are shown, at similar concentration of activity in the phantom background (3.5 kBq/ml in 3D and 3.7 kBq/ml in 2D acquisition). In 3D mode, lesion detectability resulted clearly higher than 2D mode, because of the lower noise present in 3D images, as shown in Fig. 1a. The RC values calculated in the same images are shown in Fig. 1c.
60%
40%
2D - OSEM 2 iter. x 20 subs.
20%
3D - VUE-Point 2 iter. x 20 subs.
0% 0
5
10
15
20
25
30
35
internal diameter (mm) Fig. 1. (a) The background CV versus the slice number, (b) the LOBV values and (c) the RC values versus hot spheres ID (intrinsic contrast 8:1), comparing 3D and 2D acquisition mode, respectively at 7.8 and 8.3 kBq/ml activity concentration in the background. For 2D images the number of subsets was set to 14 in order to reduce the CV near the values measured in 3D reconstructed images.
3366
A. Ferretti et al. / European Journal of Radiology 81 (2012) 3363–3370
3.2. Effect of the reconstruction parameters
3.3. Effect of phantom dimension
(a) 3D - VUE-Point 2 iter. x 14 subs.
50%
3D - VUE-Point 2 iter. x 20 subs.
background CV
3D - VUE-Point 2 iter. x 28 subs.
40%
30%
20%
10%
0% 0
5
10
15
20
25
30
35
40
45
15.6
19.7
24.8
31.3
slice position 60
(b)
3D - VUE-Point 2 iter. x 14 subs. 3D - VUE-Point 2 iter. x 20 subs.
50
3D - VUE-Point 2 iter. x 28 subs.
LOBV
40
30
20
10
0 3.9
4.9
6.2
7.8
9.8
12.4
internal diameter (mm) 120%
(c) 100%
80%
RC
According to the references [19,20], the ordered-subset iterative reconstruction reaches the convergence for large values of the product of subsets by iterations (product greater or equal to 50). It was already shown [16] that a reduction in this product generates images with a low noise level but at the expense of a biased contrast, in particular in low-count regions. In order to reach a good compromise between the contrast accuracy and a low variance, we compared PET images, acquired in 3D mode, reconstructed with the clinically implemented method (20 subsets and 2 iterations) with the same raw data reconstructed with two different number of subsets (14 and 28). In Fig. 2a the mean CV values measured in a circular ROI (8.7 cm2 ) positioned in the central area of the phantom are shown for the three different number of subsets, while the calculated LOBV and RC values for these reconstruction modalities are shown in Fig. 2b and c for the PET/CT images with 7.8 kBq/ml in background. Increasing the product subsets by iterations, images resulted noiser, and so the LOBV values decreased, while the measured contrast (i.e. RC values) improved. In Fig. 3a the mean CV values and the measured spatial resolution in air, for different post-reconstruction filters (3, 5 and 7 mm FWHM) are shown. The increased post-processing filter of 7 mm FWHM caused a smoothing in images (i.e. worsening in spatial resolution), while it reduced the noise variance (i.e. CV values). The impact on PVE is shown in Fig. 3b and c. As regards LOBV values (Fig. 3b), stronger filter caused an improvement in lesion detectability, particularly evident passing from 3 to 5 mm FWHM filter. Conversely, the contrast underestimation (Fig. 3c) had an inverse trend, worsening when the filter grew. In order to test the combined effect of the number of iterations and the post-reconstruction smoothing, we compared the subsequent cases: (i) 14 subsets × 2 iterations and 3 mm FWHM smoothing, (ii) 20 subsets × 2 iterations and 5 mm smoothing, (iii) 28 subsets × 2 iterations and 7 mm smoothing, as reported in Fig. 4. Finally, the impact of the reconstruction matrix on PVE is shown in Fig. 5a–c. As CV values did not change significantly from 128 × 128 to 256 × 256 reconstruction matrix (Fig. 5a), the lesion detectability was quite constant for the bigger spheres (Fig. 5b). A relevant increase in measured contrast was recorded with the finest matrix for spheres ID under 2 cm, together with an increased lesion detectability (Fig. 5c). Also the Nyquist sampling theorem substantiated this findings. According to it, the optimal pixel size is equal to half the spatial resolution. A greater pixel size produces undersampling and so aliasing, while smaller pixel size produces useless oversampling. Thus, the 256 × 256 matrix, with pixel size of 2.7 mm coupled to our post-reconstruction filter at 5 mm FWHM, resulted in the optimal choice.
60%
60%
40% 3D - VUE-Point 2 iter. x 14 subs.
A further aspect investigated regards the patient dimensions. In order to evaluate PVE dependence on absorber thickness, we performed further acquisitions inserting our cylindrical phantom in a water tank, increasing the water thickness of 10 cm laterally and of 7 cm vertically. The additional scattering and attenuation of photons produced a slight increase in noise (see Fig. 6a), and consequently lower values of LOBV (Fig. 6b). Conversely, the RC values did not changed significantly, as shown in Fig. 6c. 3.4. The low-dose acquisition protocol In our institution a low dose protocol was adopted for 18 F-FDG PET/CT exams, injecting to each patient 2.2 MBq/kg, in order to
3D - VUE-Point 2 iter. x 20 subs.
20%
3D - VUE-Point 2 iter. x 28 subs.
0% 0
10
20
30
40
internal diameter (mm) Fig. 2. (a) The mean CV, (b) the LOBV values and (c) the RC values versus spheres internal diameters (intrinsic contrast 8:1), comparing 3D data acquired at an activity concentration of 8.3 kBq/ml and reconstructed by VUE-Point algorithm with 2 iterations but different number of subsets (14, 20 and 28).
A. Ferretti et al. / European Journal of Radiology 81 (2012) 3363–3370
12%
12
3367
120%
(a) mean CV (left axis) FWHM (right axis)
100%
10 10%
8%
RC
FWHM
mean CV
80% 8
60%
6
40% 6% 4 3D - VUE-Point 2x14 filter 3mm 3D - VUE-Point 2x20 filter 5mm 3D - VUE VUE-Poin Pointt 2x2 2x28 8 fil filter ter 7m 7mm m
20% 4%
2 3mm
5mm
7mm
0%
post-reconstruction filter
0
(b) 50
3D - VUE-Point filter 3mm 3D - VUE-Point filter 5mm
20
30
40
Fig. 4. The RC values versus spheres ID (intrinsic contrast 8:1) comparing 3D data acquired at an activity concentration of 8.3 kBq/ml and reconstructed with different combinations of iterations and smoothing.
3D - VUE-Point filter 7mm
40
LOBV
10
internal diameter (mm)
60
decrease patient internal exposure. CT and PET images are acquired 1 h later, with an activity reduced to 1.5 MBq/kg (equivalent to 1.5 kBq/ml in our cylindrical phantom). To quantify the small lesion detectability and the SUV accuracy for this protocol, we acquired PET scans of the same Jaszczak phantom exploiting the 18 F isotope decay, decreasing progressively the concentration of radioactivity and consequently the count rate. Spheres to background intrinsic contrast was kept always equal to 8:1. The LOBV values versus spheres ID, obtained decreasing the activity, are reported in Fig. 7a for the six smaller spheres acquired in 3D mode at different activity concentrations of the background. Clearly the LOBV values increased with the activity (that is with counts rate), but the lesions can be distinguish from the background (LOBV greater than 1) starting from a diameter of 6.2 mm. In addition, Fig. 7b confirms that adopting this low dose protocol, the SUV accuracy does not deteriorate. In fact RC versus spheres diameters followed the same curve at different activities.
30
20
10
0 3.9
4.9
6.2
7.8
9.8
12.4
15.6
19.7
24.8
31.3
internal diameter (mm) 120%
(c) 100%
4. Discussion
RC
80%
60%
40% 3D - VUE-Point filter 3mm 3D - VUE-Point filter 5mm
20%
3D - VUE-Point filter 7mm
0% 0
10
20
30
40
internal diameter (mm) Fig. 3. (a) The mean CV and the spatial resolution measure in air, (b) the LOBV values and (c) the RC values versus spheres ID (intrinsic contrast 8:1), comparing 3D data acquired at an activity concentration of 7.8 kBq/ml and reconstructed applying three different smoothing filter (3, 5 and 7 mm FWHM).
Many studies were performed to evaluate the PVE in PET/CT scanners, both for 2D and 3D acquisition modalities. Usually those measurements were performed with an elliptical body phantom equipped with hollow spheres with ID ranging from 10 to 37 mm [4–7]. Some authors also measured the SUV underestimation in PET images by means of smaller spheres [8,9]. Also the impact of intrinsic contrast of lesions over background, image noise and spatial resolution on SUV accuracy was calculated by computer simulations [9,10]. Aim of the present study was to optimize the acquisition and reconstruction protocols clinically used with patients in our center, in order to maximize the SUV accuracy and the detectability of lesions with dimensions near the PET spatial resolution. A good compromise between PET quantitative accuracy and noise variance (quantities in competition during the iterative reconstruction process) was reached acquiring PET/CT images in 3D mode and reconstructing them with VUE-Point algorithm (a fully-3D OSEMtype iterative method). We investigated the impact on PVE of the reconstruction process, in particular as regards: (a) the overall iterations used (result of the product between subsets number
3368
A. Ferretti et al. / European Journal of Radiology 81 (2012) 3363–3370
60%
60%
(a)
(a)
phantom only
128x128 3D - VUE-Point 256x256 3D - VUE-Point
50%
background CV
40%
background CV
phantom with additional thickness
50%
30%
40%
30%
20%
20%
10%
10%
0%
0% 0
5
10
15
20
25
30
35
40
5
0
45
10
50
45 3D - VUE-Point 256x256
40
25
30
35
40
45
15.6
19.7
24.8
31.3
phantom only
3D - VUE-Point 128x128
40
20
(b)
(b) 45
15
slice position
slice position
phantom with additional thickness
35
35
30
LOBV
LOBV
30 25
25 20
20 15
15 10
10
5
5
0
0 3.9
4.9
6.2
7.8
9.8
12.4
15.6
19.7
24.8
3.9
31.3
4.9
6.2
9.8
12.4
120%
120%
(c)
(c) 100%
100%
80%
RC
80%
RC
7.8
internal diameter (mm)
internal diameter (mm)
60%
60%
40%
40%
phantom only
20%
3D - VUE-Point - 256x256
20%
phantom with additional thickness
3D - VUE-Point - 128x128
0% 0%
0 0
10
20
30
40
internal diameter (mm) Fig. 5. (a) The background CV versus slice position, (b) the LOBV and (c) RC values versus spheres ID, comparing two reconstruction matrixes (128 × 128 and 256 × 256 pixel), for 3D VUE-Point reconstructions at 7.8 kBq/ml activity concentration in the background and 8:1 hot spheres intrinsic contrast.
10
20
30
40
internal diameter (mm) Fig. 6. (a) The background CV versus slice position, (b) the LOBV and (c) RC values versus spheres ID, with and without an additional absorber (inserting the cylindrical phantom in a water tank 30 cm × 28 cm × 35 cm), acquired in 3D mode at 7.8 kBq/ml activity concentration in the background and 8:1 hot spheres intrinsic contrast.
A. Ferretti et al. / European Journal of Radiology 81 (2012) 3363–3370
20
The limit of lesion detectability was identified about 6–7 mm lesion diameter, also with the low-activity protocol adopted in our institute. This low-dose protocol (2.2 MBq/kg b.w. 18 F-FDG activity injected to patients) assured a low internal exposition to the patient and an unchanged SUV accuracy, as shown by the present measurements. The whole-body PET/CT image quality was evaluated by experienced nuclear medicine physicians in clinical images and it resulted adequate, allowing to keep restrained the exam duration (about 20 min for a whole-body scan) and reducing the internal exposition of patients. Furthermore, we tested how the quantitative accuracy changed with patients thickness, repeating the measurements inserting the phantom in an additional container (30 cm × 28 cm × 35 cm) filled with water. We found no significantly effects on SUV accuracy (i.e. RC values), but the small lesion detectability over the background was reduced. Clinically, these data suggested some further caution in the interpretation of PET/CT images in obese patients due to a slight increase of the noise. An issue that was not treated in this work regards the implementation of the PVE correction, that requires further phantom measurements for different values of intrinsic lesion-to-background contrast. This aspect remains an open issue, to be developed soon.
(a) 3D - 1.5 kBq/ml - 3 min 3D - 3.5 kBq/ml - 3 min 3D - 7.8 kBq/ml - 3 min
LOBV
15
10
5
0 3.9
4.9
6.2
7.8
9.8
3369
12.4
internal diameter (mm) 120%
(b) 100%
5. Conclusions
RC
80%
60%
40% 3D - 1.5 kBq/ml - 3 min 3D - 3.5 kBq/ ml - 3 min
20%
3D - 7.8 kBq/ml - 3 min
0%
0
10
20
30
40
internal diameter (mm) Fig. 7. (a) The LOBV values of the smallest hot spheres and (b) the RC values versus spheres internal diameter, for images acquired at three different values of activity concentration in the background (3D VUE-Point reconstruction).
and iterations number) and (b) the post-processing filter used to reduce the image noise. As regards the first issue, in literature the debate about the minimum number of overall iterations necessary to assure the iteration algorithm convergence is already open [16]. We chose to measure the effect comparing three cases, reconstructing the same PET raw data with 14, 20 and 28 subsets and 2 iterations, respectively corresponding to a total of 28, 40 and 56 iterations in the expectation–maximization method. Decreasing the number of subsets, the noise variance decrease, increasing the lesion detectability (i.e. LOBV values). Conversely, the RC curve versus spheres ID became less steep and so the measured contrast worsen for lesions ranging from 8 to 20 mm. As regards the second issue, the post-processing filter is used to reduce the image noise produced by the iterative algorithm. An additional smoothing, coupled with a high number of iterations, could reduce the noise variance, but producing a biased lesion-to-background contrast. We tested also the possibility to vary both the number of iterations and the smoothing filter, concluding that images obtained with 20 subsets × 2 iterations and 5 mm FWHM postreconstruction filter reached a good compromise between lesion detectability over the background and contrast accuracy.
The PET/CT scanner GE Discovery STE showed good performances, mainly in 3D acquisition modality, as regards both the detection of micro-volume hot spheres and the noise variance. The data shown in this work were collected using a set of very small hollow spheres, even smaller than the resolution of the PET scanner, with a sphere to background intrinsic contrast of 8:1, similar to clinical SUV values found in patients. Present measurements showed that hot spheres greater than 6 mm resulted detectable also simulating the low dose protocol used in clinical cases (2.2 MBq/kg injected 1 h before scan start). The partial volume effect was quantified by means of the RC. These values decreased under 90% for spheres ID smaller or equal to 12.4 mm, and it is about 50% for spheres internal diameter of 10 mm. RC values appeared to be not dependent on the concentration of activity present in the spheres at fixed scan duration (i.e. the total counts acquired), but they are affected by the number of subsets of the iterative reconstruction and the post-processing smoothing filter used. In particular these parameters had an inverse impact on RC and LOBV: measured contrast improved while noise and LOBV values worsened when the number of subsets were increased or the smoothing was reduced. The standard 3D scan reconstructed by VUE-Point algorithm with 20 subsets × 2 iterations and 5 mm FWHM filter appeared to be a good compromise between low noise variance and high SUV accuracy in small lesions. Even if the scatter and random coincidences acquired in 2D mode are less than in 3D mode, the improved VUE-Point iterative algorithm available in our hybrid scanner for fully-3D reconstructions produces PET images with a very low noise level and a good quantitative accuracy, with a final image quality higher than 2D-OSEM PET images produced by the same system.
References [1] Townsend DW. Positron emission tomography/computed tomography. Seminars in Nuclear Medicine 2008;38:152–66. [2] Sattler B, Lee JA, Lonsdale M, Coche E. PET/CT (and CT) instrumentation, image reconstruction and data transfer for radiotherapy planning. Radiotherapy and Oncology 2010;96:288–97. [3] Weber WA. Quantitative analysis of PET studies. Radiotherapy and Oncology 2010;96:308–10. [4] Geworski L, Knoop BO, Levi de Cabreias M, Knapp WH, Munz DL. Recovery correction for quantitation in emission tomography: a feasibility study. European Journal of Nuclear Medicine 2000;27:161–9.
3370
A. Ferretti et al. / European Journal of Radiology 81 (2012) 3363–3370
[5] Soret M, Bacharach SL, Buvat I. Partial-volume effect in PET tumor imaging. Journal of Nuclear Medicine 2007;48:932–45. [6] Brown C, Dempsey MF, Gilen G, Elliott AT. Investigation of 18 F-FDG 3D mode PET image quality versus acquisition time. Nuclear Medicine Communications 2001;31:254–9. [7] Bettinardi V, Mancosu P, Danna M, et al. Two-dimensional vs three-dimensional imaging in whole body oncologic PET/CT: a Discovery STE phantom and patient study. Quarterly Journal of Nuclear Medicine and Molecular Imaging 2007;51:214–23. [8] Raylman RR, Kison PV, Wahl RL. Capabilities of two- and three-dimensional FDG-PET for detecting small lesions and lymph nodes in the upper torso: a dynamic phantom study. European Journal of Nuclear Medicine 1999;26:39–45. [9] Brambilla M, Matheoud R, Secco C, et al. Impact of target-to-background ratio, target size, emission scan duration and activity on physical figures of merit for a 3D LSO-based whole body PET/CT scanner. Medical Physics 2007;34:3854–65. [10] Boellaard R, Krak NC, Hoekstra OS, Lammertsma AA. Effect of noise, image resolution, and ROI definition on the accuracy of standard uptake values: a simulation study. Journal of Nuclear Medicine 2004;45:1519–27. [11] Teo BK, Seo Y, Bacharach SL, et al. Partial-volume correction in PET: validation of an iterative postreconstruction method with phantom and patient data. Journal of Nuclear Medicine 2007;48:802–10. [12] Srinivas SM, Dhurairaj T, Basu S, Bural G, Surti S, Alavi A. A recovery coefficient method for partial volume correction of PET images. Annals of Nuclear Medicine 2009;23:341–8.
[13] Barbee1 DL, Flynn RT, Holden JE, Nickles RJ, Jeraj R. Partial volume correction of PET-imaged tumour heterogeneity using expectation maximization with a spatially varying point spread function. Physics in Medicine & Biology 2010;55:221–36. [14] Macdonald LR, Schmitz RE, Alessio AM, et al. Measured count-rate performance of the Discovery STE PET/CT scanner in 2D, 3D and partial collimation acquisition modes. Physics in Medicine & Biology 2008;53:3723–38. [15] Teras M, Tolvanen T, Johansson JJ, Williams JJ, Knuuti J. Performance of the new generation of whole-body PET/CT scanners: Discovery STE and Discovery VCT. European Journal of Nuclear Medicine and Molecular Imaging 2007;34:1683–92. [16] Soret A, Boellaard R, van der Weerdt A. Authors of the letter and the reply – number of iterations when comparing MLEM/OSEM with FBP. Journal of Nuclear Medicine 2004;45:2125–6. [17] Burgess AE. The Rose model, revised. Journal of the Optical Society of America 1999;16:633–46. [18] International Electrotechnical Commission, IEC 61675-1, Radionuclide imaging devices. Characteristics and test conditions. Part 1: Positron emission tomographs; 1998. [19] Boellaard R, O’Doherty MJ, Weber WA, et al. EANM oncology committee and EANM physics committee, FDG PET and PET/CT: EANM procedure guidelines for tumour PET imaging: version 1.0. European Journal of Nuclear Medicine and Molecular Imaging 2010;37:181–200. [20] Hutton BF, Hudson HM, Beekman FJ. A clinical perspective of accelerated statistical reconstruction. European Journal of Nuclear Medicine 1997;24:797–808.