Hybrid scatter correction applied to quantitativeholmium-166 SPECT
Tim C de Wit1,2, Jianbin Xiao1,2, J Frank W Nijsen1,Fred D van het Schip1, Steven G Staelens1,3, Peter P van Rijk1 andFreek J Beekman1,2
1 Department of Nuclear Medicine, Image Sciences Institute, University Medical Centre Utrecht,Universiteitsweg 100, STR 5.203, Utrecht, The Netherlands2 Department of Pharmacology and Anatomy, Rudolf Magnus Institute of Neuroscience,University Medical Centre Utrecht, Utrecht, The Netherlands3 ELIS Department, Ghent University Sint-Pietersnieuwstraat, 41 B-9000 Ghent, Belgium
Received 14 March 2006, in final form 27 July 2006Published 8 September 2006Online at stacks.iop.org/PMB/51/4773
AbstractHo-166 is a combined beta–gamma emitter of which the betas can be usedtherapeutically. From the 81 keV gammas of Ho-166, SPECT images canbe obtained, which give opportunities to guide Ho-166 therapy. Accuratereconstruction of Ho-166 images is currently hampered by photopeak-scatterin the patient, down-scatter in the detector, collimator and patient caused by the1.4 MeV photons and by bremsstrahlung. We developed and validated amethod for quantitative SPECT of Ho-166 that involves correction for bothtypes of scatter plus non-uniform attenuation correction using attenuationmaps. Photopeak-scatter (S) is compensated for by a rapid 3D Monte Carlo(MC) method that is incorporated in ordered subset (OS) reconstruction of theemission data, together with simultaneous correction for attenuation (A) anddetector response (D); this method is referred to as OS-ADS. Additionally, forcorrection of down-scatter, we use a 14 keV wide energy window centred at118 keV (OS-ADSS). Due to a limited number of available energy windows,the same 118 keV energy window was used for down-scatter correction ofthe simultaneously acquired Gd-153 transmission data. Validations wereperformed using physical phantom experiments carried out on a dual-headSPECT system; Gd-153 transmission line sources were used for acquiringattenuation maps. For quantitative comparison of OS-ADS and OS-ADSS,bottles filled with Ho-166 were placed in both a cylindrical phantom andan anthropomorphic thorax phantom. Both OS-ADS and OS-ADSS werecompared with an ordered subset reconstruction without any scatter correction(OS-AD). Underestimations of about 20% in the attenuation map were reducedto a few per cent after down-scatter correction. The average deviationfrom the true activity contained in the bottles was +72% with OS-AD.
Using OS-ADS, this average overestimation was reduced to +28% and withOS-ADSS the deviation was further reduced to 16%. With OS-AD and OS-ADS, these numbers were more sensitive to the choice of volumes of interestthan with OS-ADSS. For the reconstructed activity distributions, erroneousbackground activity found with OS-AD was reduced by a factor of ∼2 byapplying OS-ADS and reduced by a factor of ∼4 by applying OS-ADSS.The combined attenuation, photopeak-scatter and down-scatter correctionframework proposed here greatly enhanced the quantitative accuracy of Ho-166imaging, which is of the uppermost importance for image-guided therapies. Itis expected that the method, with adapted window settings, also can be appliedto other isotopes with high energy peaks that contaminate the photopeak data,such as I-131 or In-111.
1. Introduction
Recent years have shown a growing interest in the radioactive isotope Ho-166 for the treatmentof various types of cancer (Zielhuis et al 2005). Possible applications include skeletal internalradiotherapy (STR) which labels Ho-166 to a compound that specifically attaches to bone andbone marrow, delivering local radiation therapy and thereby sparing surrounding organs andtissue from side effects of treatment (Giralt et al 2003), or the treatment of liver malignanciesby injecting Ho-166-loaded microspheres (Nijsen et al 2002, Mumper et al 1991).
The usage of quantitative SPECT imaging for dose estimation, based on the 81 keVgamma emissions of Ho-166, may be very useful to optimize the balance of β-radiationdose to healthy tissue and tumours. There are some difficulties with obtaining quantitativeHo-166 SPECT images, however. Like other SPECT procedures, correction for attenuationand photopeak-scatter has to be applied to improve quantitative accuracy, preferably withmodelling the collimator response in order to get further improvement of resolutionand contrast-to-noise ratio (CNR). However, an additional and significant obstruction forquantitative SPECT reconstruction is the occurrence of multiple gamma emissions in the MeVrange (1.38 MeV/0.93%, 1.58 MeV/0.19%, 1.66 MeV/0.12%)4 of the Ho-166 spectrum,causing interactions within the patient, collimator and crystal. Furthermore, there is aconsiderable amount of bremsstrahlung, caused by β− (Emax: 1.77 Mev/49%, 1.85 MeV/50%)interactions within the patient. In this paper, ‘photopeak-scatter’ is defined as all primaryphotopeak photons that underwent one or more interactions inside the object but still end upin the photopeak window. The contribution of these photons can be accurately calculated withMonte Carlo simulation. ‘Down-scatter’ on the other hand is defined here as all photons thatend up in the emission or transmission window that originate from the MeV photons or theβ− emissions. These photons come from bremsstrahlung and from interactions of the MeVphotons inside the object, the collimator (septal penetration) and the crystal. In the following,the combination of photopeak-scatter and down-scatter will be denoted by ‘scatter’. Thedecrease of the number of primary photons in the photopeak window due to photo-electricand Compton interactions is referred to as ‘attenuation’.
For planar Anger camera imaging, a dual energy window-scatter subtraction methodhas been proposed for quantifying Ho-166 uptake (Bayouth and Macey 1994). However, theinherent disadvantage of planar imaging compared to SPECT is its relative poor performancefor quantitative analysis. This is due to the impossibility of doing accurate attenuation
4 For comparison, there is a 6.7% abundance of 80.6 keV photons.
correction and the difficulty of distinguishing underlying and overlying activity of overlappingorgans. So far, no solutions have been reported in the literature for quantitative SPECT of Ho-166 distributions. In this paper we propose a hybrid correction method for quantitative Ho-166SPECT imaging. To this end we combine attenuation correction, resolution recovery, MonteCarlo (MC)-based photopeak-scatter correction and window-based down-scatter correction.This method is validated by physical experiments with differently sized phantoms.
2. Materials and methods
By now, the application of MC simulation is generally accepted as a valuable tool forquantitative SPECT imaging (Lazaro et al 2005, Zaidi and Koral 2004, Buvat and Castiglioni2002, De Jong et al 2002, Dewaraja et al 2000a, 2000b, Ljungberg and Strand 1990, DeVries et al 1990, Ljungberg and Strand 1989, Floyd et al 1986). It is mostly used to modelthe photopeak-scatter for relatively low energy photons. Recently it even became possible todo fully 3D MC-based photopeak-scatter modelling in a statistical reconstruction framework(Beekman et al 1999), by using extremely fast MC simulators (De Jong et al 2001). Model-based photopeak-scatter correction methods, such as MC-based photopeak-scatter correction,have the potential to produce images with improved contrast, resolution, quantitative accuracyand signal-to-noise ratio (SNR) compared to window-based photopeak-scatter subtractionmethods and 2D methods (Xiao et al 2005, Frey et al 2002, Beekman et al 1996, 1997).
Fast MC simulators in 3D statistical reconstruction use convolution-based forced detection(CFD) of medium and low energy photons (<200 keV) for calculation speed-up, but are notsuitable yet for modelling collimator and crystal interactions of high energy photons such asthe MeV radiation. The down-scatter projections, however, are smooth and do not containmuch spatial information. Therefore a hybrid method using a window-based approach fordown-scatter correction and MC-based photopeak-scatter correction is a practical methodwhich will be developed in the present paper.
2.1. Data acquisition
All phantom measurements (section 2.3) were performed on a dual-head ADAC Vertex MCDcamera equipped with medium energy (MEGP) collimators and a Gd-153 scanning linesource (5.5 GBq) for transmission CT. The SPECT system resolution measured according tothe NEMA protocol was 1.0 cm (1.6 cm) full width at half maximum (FWHM) for a Ho-166point source at 10 cm (20 cm) distance from the MEGP collimator.
The Ho-166 photopeak window was set to 81 keV ±7.5%. Transmission data werecollected simultaneously in a 15% wide energy window centred at 100 keV inside a 5 cmwide electronic window (Gagnon et al 1999), moving along with the transmission source. Anadditional 12% wide energy window centred at 118 keV was acquired to obtain an estimatefor the down-scatter in both the emission and transmission window.
A 360◦ study for simultaneously obtaining transmission and emission data was acquiredover 120 projections (60 projections per head). The acquisition time was 700 s per projectionfor the emission scan and 29 s for the transmission scan. Frame-by-frame decay correction wasperformed afterwards. The acquisition matrix was 128 × 128 with a pixel size of 4.72 mmfor both emission and transmission data. In order to assess the influence of noise on thequantitative results, ten clinically realistic noise realizations, corresponding to 30 s perprojection, were emulated by adding Poisson noise to the high-count emission data, afterdown-scaling.
Table 1. Overview of the emission reconstruction methods OS-AD, OS-ADS and OS-ADSS.For each method we show the factors that are corrected for in the forward projection and backprojection: attenuation (A), detector blurring (D), MC-based photopeak-scatter correction (SMC)and down-scatter correction by using the 118 keV energy window (S118). In the right column welist the energy windows that were used. For all methods, an additional 100 keV window was usedfor acquiring transmission data.
Reconstruction Reconstruction Forward projection Back projection Energy
method algorithm A D SMC S118 A D SMC S118 windows
As the SPECT system did not allow for intrinsic flood corrections for each energy windowseparately, additional extrinsic flood correction was necessary for all energy windows toeliminate non-uniformity effects in the projections.
2.2. Reconstruction methods
2.2.1. Iterative emission reconstruction. Fully 3D MC-based SPECT image reconstructionwas performed using the ordered subsets expectation maximization (OSEM) algorithm(Hudson and Larkin 1994, Lange and Carson 1984, Shepp and Vardi 1982). OSEM withattenuation and detector blurring modelling within both forward and back projection isreferred to as OS-AD. OS-AD with additional modelling of photopeak-scatter within theforward projection is called OS-ADS. Since for OS-ADS the forward projector is not identicalto the back projector this method is referred to as dual matrix ordered subsets expectationmaximization (DM-OSEM) (Kamphuis et al 1998, Kadrmas et al 1998). An accelerated MCsimulator (Utrecht Monte Carlo system) was used to perform detailed modelling of photopeak-scatter in the forward projection part of the DM-OSEM algorithm (Beekman et al 2002). Dueto the implementation of variance reduction techniques such as convolution-based forceddetection (De Jong et al 2001), 105 photon tracks per subset were sufficient to produce lownoise photopeak-scatter projections (de Wit et al 2005). By taking the measured down-scatterdata into account during reconstruction, the images were further improved, compared toOS-ADS; this reconstruction method is referred to as OS-ADSS. In table 1 we show anoverview of all three emission reconstruction methods.
A proof of concept of the OS-ADSS method will be demonstrated in the next sub-section,using planar point source measurements. To illustrate how the down-scatter is taken intoaccount, here we show the DM-OSEM algorithm with a background term included
λn+1i (k) = λn
i (k)∑j∈Sp(n) bij
∑j∈Sp(n)
bijpj∑i cij λ
ni (k) + Bj
(1)
where λn+1i (k) represents the ith source voxel of the updated iteration after processing subset
n, Sp(n) contains the projection angles of subset n, and k represents the iteration number. Thematrix elements used for forward projection are given by cij and for back projection by bij .The measured projections for bin j are given by pj . The background term Bj represents anestimate of the down-scatter contribution in the photopeak window, which is derived from the118 keV window according to equation (3). Equation (1) is the dual matrix OS version ofthe MLEM method proposed by King et al (1997) and Bowsher et al (1996) to correct forscatter and background radiation. For all three reconstruction methods (OS-AD, OS-ADS,
Figure 1. Raw down-scatter data from the 118 keV energy window (left) and maximum likelihoodde-noised down-scatter data (right), with associated horizontal profiles.
OS-ADSS) a relatively high number of iterations (40) with eight subsets were used, sinceseveral studies have shown that better results can be obtained by using over-iteration and post-filtering, than with early stopping of the iteration process (Nuyts and Fessler 2003, Beekmanet al 1998, Miller and Wallis 1992). At this combination of iterations and subsets a fairlyuniform resolution can be obtained when detector blurring is modelled.
In the reconstruction of the activity distributions, simultaneously acquired andreconstructed attenuation maps were used to facilitate non-uniform attenuation correctionand anatomy-dependent photopeak-scatter correction. The attenuation values for each photonenergy were calculated by multiplying the density map by the attenuation coefficient of waterfor that energy, thereby assuming a water equivalent density distribution (Beekman et al 2002).
2.2.2. Iterative transmission reconstruction. Attenuation maps were reconstructed using theordered subset convex (OSC) (Kamphuis and Beekman 1998) algorithm with eight subsets and30 iterations. In order to reduce noise, the attenuation maps were 3D Gaussian filtered (sigma5.2 mm) followed by 3D cubic median filtering twice (kernel size 14.2 × 14.2 × 14.2 mm3).Similar to OS-ADSS emission reconstruction, down-scatter can be taken into account in theOSC algorithm. The OSC algorithm updates the attenuation coefficients µn
s (k), for each voxelk in the object, for subset s and iteration n, according to
µns+1(k) = µn
s (k) + µns (k)
∑i∈S(n) lik
(1 − Yi
yi(µns )+Bi
)yi
(µn
s
)∑
i∈S(n) likyi
(µn
s
)pi
(µn
s
) (2)
where lik is the length of the projection line i through voxel j, S(n) contains the projectionsin subset s, Yi is the measured number of photons at detector bin i and yi
(µn
s
)is the expected
number of photons at bin i, yi
(µn
s
) = di exp(−pi
(µn
s
)), which is a function of the blank
scan counts di in bin i, and the forward projection pi
(µn
s
)of the attenuation distribution at
detector bin i, given by pi
(µn
s
) = ∑j lijµ
ns (j). The background term Bi is an estimate of
the down-scatter contribution in the transmission window which is derived from the 118 keVwindow according to equation (3).
2.2.3. Window-based down-scatter correction. A down-scatter estimate for both the emissionand transmission window was obtained in the 118 keV energy window. The noise in the down-scatter data was suppressed using maximum likelihood fitting (60 iterations) of Gaussian-shaped basis functions (sigma = 12 mm) using the method described in Colijn and Beekman(2004). The resulting noise reduction of the projections, by taking this step, is illustrated infigure 1.
The background-term in (1) and (2) can be written as
Bi = KQi (3)
where Qi is the de-noised down-scatter data for bin i and K is a scaling factor. This implies thatduring reconstruction a fraction K of the down-scatter data is added to the forward projector
Figure 2. PSF measurements in the 100 keV energy window (solid profiles), reflecting the truedown-scatter into the transmission window. The dotted profiles were taken from a down-scatterwindow at 118 keV, indicating that the 118 keV down-scatter responses provide an accurate estimatefor down-scatter correction of the transmission data.
of the maximum likelihood algorithm. For emission reconstruction, the appropriate scalingfactor, KEM, equalled the ratio of the width of the down-scatter window over the width of thephotopeak window (KEM = 12.15/14.16 = 0.858). For transmission reconstruction a scalingfactor KTR was used, which consists of two parts
KTR = KTR,ww KTR,acq (4)
where KTR,ww equals the ratio of the width of the transmission window over the width of thedown-scatter window (KTR,ww = 15.00/14.16 = 1.06) and KTR,acq depends on the acquisitionparameters (the acquisition times for emission and transmission, and the width of the movingelectronic window). The latter factor was calculated based on the method described in Zhaoet al (2005).
2.2.4. Down-scatter characteristics. In order to validate the scaling factor as obtained bycalculation based on energy window widths, planar images of a Ho-166 point source in air,35 cm from the collimator were recorded. The measurement was repeated, with a 15 cm slabof Perspex between the point source and the collimator. Responses were measured in the100 keV energy window (true down-scatter in the transmission window) and in the 118 keVenergy window (down-scatter estimate). The horizontal profiles corresponding to these point-source responses are shown in figure 2, together with down-scatter profiles, scaled accordingto equation (4). This shows that regardless of the source-detector distance and scatter medium,most of the down-scatter tails can be successfully removed by using the pre-calculated scalingfactor KTR,ww.
Down-scatter correction can also be used for the photopeak window (81 keV), usingthe same 118 keV down-scatter window. However, the effectiveness of the down-scattercorrection cannot be clearly illustrated in the same way as above, due to a fair amount ofadditional photopeak-scatter for the 81 keV photons. In this case it is more instructive tolook at the energy spectrum, as is shown in figure 3. The solid spectrum was measuredfrom a point source in air. The dotted spectrum was measured from a point source inside awater-filled thorax phantom, showing the influence of photopeak-scatter that is corrected inOS-ADS and OS-ADSS using MC modelling; more importantly, it shows that the propertiesof the down-scatter photons above 100 keV are highly resemblant for both cases due to thehigh energy of the originating photons. Furthermore, the ratio between the photopeak and
Quantitative SPECT with holmium 4779
Figure 3. The energy spectrum of Ho-166 for a point source in air and for a Ho-166 distribution ina thorax phantom, normalized to the background above 100 keV for easy comparison. The threeenergy windows that are used are represented by grey bars. These spectra clearly illustrate thatboth 81 keV and 100 keV data are severely contaminated with down-scattered photons and thatthe energy window at 118 keV can be used for down-scatter correction, both for the 81 keV and100 keV energy window.
the background level is constant. A simple calculation shows that the backscatter peak andCompton edge for the 1.4 MeV peak are located at 0.22 MeV and 1.2 MeV, respectively. Thismeans it is very unlikely for low-order Compton-scattered MeV photons to be detected in theemission or transmission window. The detected down-scatter distribution therefore will notcontain much spatial information, leading to K-factors that are rather independent of objectsize/shape. Therefore, it was assumed that the 118 keV energy window is also suitable fordown-scatter correction in the 81 keV photopeak data. Acquiring the down-scatter estimatefor both emission and transmission in a single energy window was necessary due to the limitednumber (three) of energy windows available.
2.2.5. Projection masking. The significant amount of down-scatter of holmium, even faroutside the detector area where the object is projected, gives rise to an additional background inthe reconstruction. This often leads to reconstruction artefacts since the algorithm may ‘expect’a very low amount of counts outside this area, based on the imperfect photon transport modelused. As a result, voxels activities along the object borders (defined by the attenuation map)are increased to compensate for this down-scatter, and blow up after several iterations.
These effects can be easily corrected for by using a scheme to mask out everything outsidethe object area in the projections; for this we need simultaneously acquired transmission data Iand the corresponding blank scan I0. For each projection pixel i, the attenuation integral alongthe lines of response is given by∫
µ dxi = −ln
(Ii
(I0)i
).
The projections created this way are thresholded at 0.30, and are expanded by amorphological transformation (dilation) in order to prevent boundary inconsistencies. Thethreshold value was tuned such that it provided a mask that narrowly enclosed the whole
4780 Tim C de Wit et al
object, without producing ‘holes’ near the lung area in the projections. Next, the emissionprojections are multiplied by the resulting projection mask to clear the area outside the object.Complete clearing has the advantage that the MLEM or OSEM algorithm neglects these zeropixel values. The down-scatter projections are masked similarly.
2.3. Experimental validations
Two phantoms were used for validations: a cylindrical phantom (inner diameter 21.6 cm, innerheight 18.6 cm) and an anthropomorphic thorax phantom (Model ECT/TOR/P), includingair-filled lungs, a spine and a liver; both manufactured by Data Spectrum Corp., Hillsborough,NC. In both phantoms three plastic bottles (220 ml; height 11.2 cm, diameter 5.80 cm) werefixated at various locations by using a water-resistant duct tape, at least several centimetresseparated from each other to eliminate cross-talk caused by partial volume effects. For thethorax phantom, bottle 1 was located horizontally at the left-lateral side on the bottom-platenext to the liver; bottle 2 vertically between the lungs, attached to the spine; and bottle 3vertically in the middle against the chest. For the cylinder, bottle 1 was located verticallyagainst the side, on the bottom-plate; bottle 2 vertically in the centre, attached to the top-plate;and bottle 3 horizontally on the bottom plate.
For the thorax phantom the bottles were filled with 677 kBq ml−1 of holmium chloride(HoCl3) each; for the cylindrical phantom they were filled with 650 kBq ml−1 of the samesolution. For both phantoms, the background consisted of non-radioactive water.
2.4. Data analysis
The accuracy of the attenuation maps with and without down-scatter correction was comparedby drawing profiles through the reconstruction of the uniform cylinder phantom and theanthropomorphic phantom. In addition, densities measured in different regions of interestswere compared.
The activity distribution, as reconstructed by three different methods, was compared for theuniformly filled cylinder and for the thorax phantom with bottles. The reconstructions wereconverted to activity concentration maps (MBq/voxel) by scaling by the absolute detectorsensitivity, which was determined by reconstructing a Tc-99m point source with a knownactivity, using the OS-ADS method. The factor was calculated by taking into account thephotopeak abundances of Ho-166 and Tc-99m, together with the relative detector efficienciesfor both photopeak energies, which are known (Cherry et al 2003) for several NaI crystalthicknesses (5/8 inch in our case).
For the quantitative analysis, volumes of interest (VOI) were automatically generated bythresholding a smoothed OS-ADSS reconstruction, such that the three VOIs had the correcta priori known volume (220 ml) of the bottle. To investigate the sensitivity of OS-AD, OS-ADS and OS-ADSS to the choice of VOI, additional VOI-sets were constructed, based on amorphological transformation (dilation) of the original VOI, expanding them by one, two andthree pixels in all directions. The activity counts were calculated using these region masks,for the unfiltered, over-iterated images, where the spill-over is minimal. Furthermore, forboth phantoms two VOIs were drawn in a background region where there is no activity in thephantom.
3. Results
Figures 4–6 show the attenuation map results for the uniform cylindrical phantom and theanthropomorphic thorax phantom.
Quantitative SPECT with holmium 4781
uncorrected corrected
µ = 0.89 × µ water µ = 0.98 × µ water
Figure 4. Slices of the attenuation map for the cylindrical phantom with and without down-scattercorrection. For both images, the average attenuation coefficient is calculated for the rectangularregions of interest, and horizontal profiles (four pixels wide) are shown.
µ = 0.79 × µ water
uncorrected corrected
µ = 0.99 × µ water
Figure 5. Like figure 4 but for a uniform area of the thorax phantom.
uncorrected
µ = 0.79 × µ water
corrected
µ = 0.99 × µ water
Figure 6. Like figure 5 but for a lung slice.
Differences between the various reconstruction methods are depicted in figure 7, showinga slice through the cylindrical phantom (figure 7(a)) and thorax phantom (figure 7(b)) withthree bottles.
In tables 2 and 3 the quantitative results are given for the bottles placed within theuniform cylinder and within the anthropomorphic thorax phantom. For each VOI selectionmethod, the deviations of the measured activities from the true activities are given for different
4782 Tim C de Wit et al
Table 2. Deviations (%) of measured activities for the cylinder phantom. The activities weremeasured for different reconstruction methods: OS-AD, OS-ADS and OS-ADSS, and for fourdifferent VOI selections.
VOI selection Measured Measured Measuredmethod Bottle OS-AD OS-ADS OS-ADSS
Table 3. Deviations (%) of measured activities for the thorax phantom. The activities weremeasured for different reconstruction methods: OS-AD, OS-ADS and OS-ADSS, and for fourdifferent VOI selections.
VOI selection Measured Measured Measuredmethod Bottle OS-AD OS-ADS OS-ADSS
reconstruction methods (% deviation = 100 × (measured-true)/true). The true bottle activitieswere 143 MBq and 149 MBq for the cylinder phantom and thorax phantom, respectively. Theaverage deviation from the true activity contained in the bottles, averaged over all values foundin tables 2 and 3, was +72% with OS-AD. Using OS-ADS, this average overestimation wasreduced to +28% and with OS-ADSS the deviation was further reduced to 16%.
The apparent background concentrations were shown in table 4 for two differentbackground regions.
A similar analysis like in tables 2 and 3 was also performed for the ten noise realizations,for each phantom. The variation between the measured activities was well below 1%; theaverage itself agreed perfectly with the high-count results.
4. Discussion
In this paper, a hybrid reconstruction method called OS-ADSS was developed and evaluated forquantitative Ho-166 SPECT imaging. This reconstruction method consists of a combination
Quantitative SPECT with holmium 4783
(a)
(b)
(i)
(ii)
(iii)
(i)
(ii)
(iii)
Voxel
Voxel
Act
ivit
y co
nce
ntr
atio
n (
MB
q/v
oxe
l)A
ctiv
ity
con
cen
trat
ion
(M
Bq
/vo
xel)
Figure 7. Image slice showing two bottles for the cylindrical phantom (a) and thorax phantom (b),for various reconstruction methods: (i) correction for detector blurring and attenuation (OS-AD),(ii) like (i) but with photopeak-scatter modelling from MC (OS-ADS), (iii) like (ii) but with down-scatter correction (OS-ADSS). To the right we show the horizontal profiles (three pixels wide),for the regions indicated on the image slices. These image profiles show a significant differencebetween OS-AD, OS-ADS and OS-ADSS. In particular, note the reduced background activity forthe OS-ADSS case.
of model-based photopeak-scatter correction, using a fast MC simulator, and window-baseddown-scatter correction. Evaluation was performed with point-source measurements andphysical phantom experiments.
4784 Tim C de Wit et al
Table 4. Background-activity concentrations, as fractions of the true bottle activity concentrations,for different reconstruction methods: OS-AD, OS-ADS and OS-ADSS. The background VOIs(∼3000 voxels each) were drawn in regions where there should be no activity. Central slices ofthe corresponding VOI were shown in the figures on the left.
Cylinder OS-AD OS-ADS OS-ADSS
Background region 1 9.3% 5.3% 2.2%Background region 2 8.9% 4.7% 2.0%
ThoraxBackground region 1 10% 5.3% 2.5%Background region 2 6.7% 4.2% 1.8%Average 8.7% 4.9% 2.1%
From our results, the OS-ADSS reconstruction method showed a quantitativeimprovement compared to the methods OS-AD and OS-ADS. However, the quantitativeanalysis did not show a big difference between OS-ADS and OS-ADSS. For the coldbackground regions the difference was much larger; the apparent background activity forOS-ADSS was about two times lower than for OS-ADS and about four times lower thanfor OS-AD. For oncological applications this clearly will have a big influence on tumour-to-background ratios and dosimetry of healthy tissues. This large difference in the apparentbackground concentration also reflects in the quantitative results of tables 2 and 3. Here, theOS-ADSS reconstruction method is significantly less sensitive to the chosen VOI-selectionthan the OS-ADS method.
The choice to take the same down-scatter energy window for both emission andtransmission reconstruction was solely due to the limited number of energy windows available(3). However, even with only three energy windows good results can be obtained, whichmakes OS-ADSS a method that could be implemented on most SPECT systems. On systemswith more than three available or overlapping energy windows, separate energy windowscould be used for down-scatter correction of emission and transmission data. For emissionreconstruction a 5% wide energy window centred around 90 keV could be used. Comparedto 118 keV, the 90 keV energy window might provide a more accurate estimate of thedown-scatter. Furthermore, it allows for correction of down-scatter in the emission windoworiginating from the Gd-153 line source, which is not possible with the 118 keV energywindow. However, the 5% wide 90 keV energy window is quite noisy and contains about 3%of the Ho-166 photopeak, which will negatively influence the quantitative results. Choosingan energy window smaller than 5% will result in even noisier down-scatter estimates. Also,an overlapping 100 keV down-scatter window, outside the moving electronic mask, wouldprobably further improve the attenuation maps.
Note that MC-based photopeak-scatter modelling was essential for this paper. Due to thepresence of an x-ray peak in the Ho-166 spectrum around 49 keV, and an additional lead x-raypeak partially overlapping with the 81 keV photopeak, the much used triple energy windowapproach (Ogawa et al 1991) cannot be used for this isotope.
For transmission reconstruction the down-scatter window was used for correcting theattenuation maps for all three methods. If such a down-scatter correction was not performedwe expect that the results would be much worse, because both photopeak-scatter correctionand attenuation correction would be compromised. Also note that attenuation maps acquiredon SPECT/CT systems will not be affected by the Ho-166 down-scatter.
The scaling factors, used for multiplying the down-scatter estimates were calculated,based on the ratio of energy window widths. The basic assumption was that the contribution
of down-scatter is comparable for the two energy windows, independent of object size/shape.In figure 3 we showed that this is only approximately true. The contribution of down-scatterin the spectrum does not seem to be strongly dependent on the scatter medium. Therefore weexpect that tuning of the scaling factors might slightly improve our results.
The amount of down-scatter counts in the measured photopeak projections was typically inthe order of 50% for our experiments, agreeing with the energy spectrum in figure 3. However,this might seem in contradiction to the image profiles in figure 7, where the photopeak-scatterseems to have a bigger impact than the down-scatter. This apparent contradiction is caused bythe large amount of septal penetration in the collimator. The spectrum displayed in figure 3is not for one specific pixel or area, but averaged over the whole detector. Photopeak photonsare mainly detected in object areas of the detector; however, down-scatter photons are alsodetected outside these areas in the projections. The amount of down-scatter counts in theobject area of the measured photopeak projections was around 12% for our experiments.
Even though the reconstructions were over-iterated to minimize partial volume effects,most results in tables 2 and 3 still gave underestimations of the activity in the bottles, whenusing small VOIs. By enlarging the regions by two pixels (9.44 mm) in all directions tocompensate for this effect, the results are much better because blurred activity around thebottles could be included. Enlarging the regions by one more pixel in all directions did notinfluence the OS-ADSS result much; however, the OS-AD and OS-ADS results are severelyinfluenced by the apparent background, which gives rise to overestimations.
The acquisition time used for the high-count emission scans (700 s/azimuth), was chosento reduce the influence of noise on the results of this ‘proof of principle’ paper. The sensitivityto noise was assessed separately by using emulated low-count emission data (30 s/azimuth).Although the high-count data were not completely noise-free, the amount of additional noisein the low-count data is negligible compared to the generated Poisson noise.
An effect that was not compensated for in this paper is lead x-rays, produced by interactionsin the collimator. Since these x-rays typically have energies in the range between 70 and90 keV, they contribute to the Ho-166 photopeak. This might explain why the activities for thenon-dilated VOIs did not all give rise to underestimations. A solution was recently proposedfor including modelling of septal penetration and lead x-rays in a reconstruction framework(Song et al 2005), which is however still computationally demanding. This issue may be atopic of future investigation.
The method developed in this paper could be extended to other isotopes with similarcontamination due to high energy gamma and beta emissions; e.g. for I-131 and In-111.Validation of the methods presented here for other isotopes is a subject of further research.
5. Conclusion
DM-OSEM reconstruction incorporating correction for attenuation, detector blurringand photopeak-scatter can provide quantitatively accurate holmium-166 SPECT images.Furthermore, our newly developed hybrid reconstruction method incorporating additionalwindow-based down-scatter correction further enhances quantitative accuracy of both thereconstructed activity distribution and the attenuation maps.
Acknowledgments
The work was sponsored by the Dutch Science and Technology Foundation (STW) by grantUGT.6069 (TdW, FN, FvhS, FB). The work of FB is supported by the Medical Counsel of the
Dutch Organization for Scientific Research by grant 917.36.335. The contents are solely theresponsibility of the authors and do not necessarily represent the official view of the sponsoringagencies.
References
Bayouth J E and Macey D J 1994 Quantitative imaging of Ho-166 with an Anger camera Phys. Med. Biol. 39 265–79Beekman F J, De Jong H W and Slijpen E T 1999 Efficient SPECT scatter calculation in non-uniform media using
correlated Monte Carlo simulation Phys. Med. Biol. 44 N183–92Beekman F J, De Jong H W and Van Geloven S 2002 Efficient fully 3D iterative SPECT reconstruction with Monte
Carlo-based scatter compensation IEEE Trans. Med. Imaging 21 867–77Beekman F J, Kamphuis C and Frey E C 1997 Scatter compensation methods in 3D iterative SPECT reconstruction:
a simulation study Phys. Med. Biol. 42 1619–32Beekman F J, Kamphuis C and Viergever M A 1996 Improved SPECT quantitation using fully three-dimensional
iterative spatially variant scatter response compensation IEEE Trans. Med. Imaging 15 491–9Beekman F J, Slijpen E T and Niessen W J 1998 Selection of task-dependent diffusion filters for the post-processing
of SPECT images Phys. Med. Biol. 43 1713–30Bowsher J E, Johnson V E, Turkington T G, Jaszczak R J, Floyd C E and Coleman R E 1996 Bayesian reconstruction
and use of anatomical a priori information for emission tomography IEEE Trans. Med. Imaging 15 673–86Buvat I and Castiglioni I 2002 Monte Carlo simulations in SPET and PET Q. J. Nucl. Med. 46 48–61Cherry S R, Sorenson J A and Phelps M E 2003 Problems in radiation detection and measurement Physics in Nuclear
Medicine 3rd edn (Philadelphia, PA: Saunders)Colijn A P and Beekman F J 2004 Accelerated simulation of cone beam x-ray scatter projections IEEE Trans. Med.
Imaging 23 584–90De Jong H W, Beekman F J, Viergever M A and Van Rijk P P 2002 Simultaneous (99m)Tc/(201)Tl dual-isotope
SPET with Monte Carlo-based down-scatter correction Eur. J. Nucl. Med. Mol. Imaging 29 1063–71De Jong H W A M, Slijpen E T P and Beekman F J 2001 Acceleration of Monte Carlo SPECT simulation using
convolution-based forced detection IEEE Trans. Med. Imaging 48 58–64De Vries D, Moore S C, Zimmerman R E, Mueller S P, Friedland B and Lanza R C 1990 Development and validation
of a Monte-Carlo simulation of photon transport in an Anger camera IEEE Trans. Med. Imaging 9 430–8de Wit T C, Xiao J and Beekman F J 2005 Monte-Carlo based statistical SPECT reconstruction: influence of number
of photon tracks IEEE Trans. Nucl. Sci. 52 1365–9Dewaraja Y K, Ljungberg M and Koral K F 2000a Accuracy of 131I tumor quantification in radioimmunotherapy
using SPECT imaging with an ultra-high-energy collimator: Monte Carlo study J. Nucl. Med. 41 1760–67Dewaraja Y K, Ljungberg M and Koral K F 2000b Characterization of scatter and penetration using Monte Carlo
simulation in 131I imaging J. Nucl. Med. 41 123–30Floyd C E Jr, Jaszczak R J, Greer K L and Coleman R E 1986 Inverse Monte Carlo as a unified reconstruction
algorithm for ECT J. Nucl. Med. 27 1577–85Frey E C, Gilland K L and Tsui B M W 2002 Application of task-based measures of image quality to optimization and
evaluation of three-dimensional reconstruction-based compensation methods in myocardial perfusion SPECTIEEE Trans. Med. Imaging 21 1040–50
Gagnon D, Tung C H, Zeng L and Hawkins W G 1999 Design and early testing of a new medium-energy transmissiondevice for attenuation correction in SPECT and PET Proc. IEEE Medical Imaging Conf. vol 3 pp 1349–53
Giralt S et al 2003 166Ho-DOTMP plus melphalan followed by peripheral blood stem cell transplantation in patientswith multiple myeloma: results of two phase 1/2 trials Blood 102 2684–91
Hudson H M and Larkin R S 1994 Accelerated image-reconstruction using ordered subsets of projection data IEEETrans. Med. Imaging 13 601–9
Kadrmas D J, Frey E C, Karimi S S and Tsui B M W 1998 Fast implementations of reconstruction-based scattercompensation in fully 3D SPECT image reconstruction Phys. Med. Biol. 43 857–73
Kamphuis C and Beekman F J 1998 Accelerated iterative transmission CT reconstruction using an ordered subsetsconvex algorithm IEEE Trans. Med. Imaging 17 1101–5
Kamphuis C, Beekman F J, Van Rijk P P and Viergever M A 1998 Dual matrix ordered subsets reconstruction foraccelerated 3D scatter compensation in single-photon emission tomography Eur. J. Nucl. Med. 25 8–18
King M A, Devries D J, Pan T S, Pretorius P H and Case J A 1997 An investigation of the filtering of TEW scatterestimates used to compensate for scatter with ordered subset reconstructions IEEE Trans. Nucl. Sci. 44 1140–5
Lange K and Carson R 1984 EM reconstruction algorithms for emission and transmission tomography J. Comput.Assist. Tomogr. 8 306–16
Lazaro D, El Bitar Z, Breton V, Hill D and Buvat I 2005 Fully 3D Monte Carlo reconstruction in SPECT: a feasibilitystudy Phys. Med. Biol. 50 3739–54
Ljungberg M and Strand S E 1990 Scatter and attenuation correction in SPECT using density maps and Monte Carlosimulated scatter functions J. Nucl. Med. 31 1560–67
Ljungberg M and Strand S E 1989 A Monte Carlo program for the simulation of scintillation camera characteristicsComput. Methods Programs Biomed. 29 257–72
Miller T R and Wallis J W 1992 Clinically important characteristics of maximum-likelihood reconstruction J. Nucl.Med. 33 1678–84
Mumper R J, Ryo U Y and Jay M 1991 Neutron-activated holmium-166-poly (L-lactic acid) microspheres: a potentialagent for the internal radiation therapy of hepatic tumors J. Nucl. Med. 32 2139–43
Nijsen J F W, van het Schip A D, Hennink W E, Rook D W, Van Rijk P P and De Klerk J M H 2002 Advances innuclear oncology: microspheres for internal radionuclide therapy of liver metastases Curr. Med. Chem. 9 73–82
Nuyts J and Fessler J A 2003 A penalized-likelihood image reconstruction method for emission tomography, comparedto postsmoothed maximum-likelihood with matched spatial resolution IEEE Trans. Med. Imaging 22 1042–52
Ogawa K, Harata Y, Ichihara T, Kubo A and Hashimoto S 1991 A practical method for position-dependent Compton-scatter correction in single photon-emission Ct IEEE Trans. Med. Imaging 10 408–12
Shepp L A and Vardi Y 1982 Maximum likelihood reconstruction for emission tomography IEEE Trans. Med. Imaging1 113–22
Song X, Segars W P, Du Y, Tsui B M W and Frey E C 2005 Fast modelling of the collimator-detector response inMonte Carlo simulation of SPECT imaging using the angular response function Phys. Med. Biol. 50 1791–804
Xiao J, de Wit T C and Beekman F J 2005 Evaluation of 3D Monte-Carlo based scatter correction for 99mTc cardiacperfusion SPECT: a phantom study J. Nucl. Med. at press
Zaidi H and Koral K F 2004 Scatter modelling and compensation in emission tomography Eur. J. Nucl. Med. Mol.Imaging 31 761–82
Zhao Z, Ye J, Coles D and Shao L 2005 Transmission attenuation map with measured downscatter correction: Nucl.Sci. Symp. Conf. Rec. 2003 IEEE 4 2923–6
Zielhuis S W, Nijsen J F, Seppenwoolde J H, Zonnenberg B A, Bakker C J, Hennink W E, Van Rijk P P and hetSchip A D 2005 Lanthanide bearing microparticulate systems for multi-modality imaging and targeted therapyof cancer Curr. Med. Chem. Anticancer Agents 5 303–13
