433 Optimisation of supervised cluster analysis for extracting reference tissue input curves in (R)-[11C]PK11195 studies
M. Yaqub1, B.N.M. Berckel1, A. Schuitemaker2, R. Hinz3, F.E. Turkheimer4, A.A. Lammertsma1 And R. Boellaard1
1Nuclear Medicine and PET Research; 2Neurology and Alzheimer Centre, VU Universite Medical Centre, Amsterdam, The Netherlands; 3Wolfson Molecular Imaging Centre, University of Manchester, Manchester; 4Neuroscience, Imperial College London, London, UK
Background and aims: (R)-[11C]PK11195 is widely used for imaging activated microglia in the brain. To date, (R)-[11C]PK11195 studies have primarily been analysed using reference tissue methods with one of the following reference tissues: cerebellum, grey matter cerebrum or automatic extraction by cluster analysis methods.1,2 The aim of the present study was to further investigate performance of a recently published supervised cluster algorithm (SVCA),1 including a new simplified reference tissue method (SRTM) with blood volume correction3 and use of a reduced number of predefined kinetic classes2 within SVCA to improve quantification.
Methods: (R)-[11C]PK11195 data was taken from young healthy controls (YC), elderly healthy controls (OC) and Alzheimer's Disease patients (AD). Cerebellum (CER) and thalamus regions were defined manually in grey and white matter. In order to extract reference tissue curves, supervised cluster analysis methods were evaluated for 6 (SVCA6)1 and 4 kinetic classes (SVCA4).2 CER input functions were also used for comparison. Kinetic analyses were performed using both SRTM with non-linear regression and a basis function implementation (RPM) of SRTM. RPM analysis was performed both with and without correction for fractional blood volume (VB) using image derived blood input curves.3 Data were also analysed using the reversible two tissue plasma input model with correction for fractional blood volume (2T4k). Thalamus was the region interest and parameters of interest were BPND for RPM/SRTM, and DVR-1 for 2T4k. DVR is the thalamus to reference tissue distribution volume of ratio. VT using 2T4k were also evaluated for reference tissues.
Results: Averages (±s.d.) of reference tissue VT showed comparable trends for CER (1.01±0.39, 0.85±0.32, 0.78±0.23; YC, OC, AD), SVCA4 (0.92±0.38, 0.68±0.26, 0.58±0.15) and SVCA6 (1.07±0.49, 0.84±0.37, 0.82±0.20) across subject groups.
Thalamus DVR-1 estimates were best using SVCA4 (0.14±0.14, 0.29±0.07, 0.44±0.32), second for CER (0.035±0.134, 0.047±0.049, 0.066±0.159) and poorest for SVCA6, showing higher values in OC (0.075±0.133) than in AD (0.008±0.167).
Without VB correction, RPM and SRTM showed similar results. Overall, SVCA4 showed more plausible results than CER and SVCA6, i.e., better distinction between YC, OC and AD (Figure 1), and better correlation with corresponding DVR-1 (R2 = 0.47). Finally, for SVCA4 and SVCA6, RPM with VB correction did not show significantly better results than without such a correction (Figure 1).
Thalamus BPND estimated using RPM, with and without a correction for VB (VB cor.) and different reference tissue input curves (CER, SVCA4 and SVCA6).
Conclusions: Supervised cluster analysis with 4 kinetic classes is the method of choice for extracting reference tissue curves in [11C](R)-PK11195 studies.
162. Evaluation of reconstruction based partial volume correction
J.E.M. Mourik, M. Lubberink, F.H.P. van Velden, A.A. Lammertsma and R. Boellaard
Nuclear Medicine & PET Research, VU University Medical Center, Amsterdam, The Netherlands
Background and aims: The limited resolution of current clinical PET scanners results in incorrect estimation of true radioactivity, which is known as the partial volume effect. Different partial volume correction (PVC) methods exist.1 One way is to correct for partial volume during reconstruction. Such a reconstruction based PVC method improves the spatial resolution by taking the point spread function of the scanner into account. The aim of present study was to assess the accuracy of reconstruction based PVC.
Methods: The NEMA NU2 image quality phantom with spheres of varying internal diameter (10, 13, 17, 22, 28, and 37 mm) was scanned on both HR+ (resolution: ∼4.1 to 7.8 mm FWHM; CTI/Siemens) and HRRT (resolution: ∼2.3 to 3.4 mm FWHM; CTI/Siemens) scanners. This phantom was used to obtain recovery (measured/true activity) of the spheres for the different reconstruction algorithms used. In addition, dynamic [11C]flumazenil data (n = 5), acquired on both scanners were used. HR+ data were reconstructed using normalization and attenuation weighted ordered subsets expectation maximization (NAW-OSEM, 4 iterations (i), 16 subsets (s)), NAW-OSEM smoothed with a 5 mm Gaussian filter (clinical standard) and a reconstruction based PVC method (PVC-OSEM, 4i, 16s).2,3 In addition, HRRT data were reconstructed using 3D ordinary Poisson (OP) OSEM (8i, 16s), OP-OSEM smoothed with a 6 mm Gaussian filter and a HRRT PVC algorithm (PVC OP-OSEM, 16i, 16s). For each reconstruction, parametric volume of distribution (VT) images were generated using a basis function implementation of the standard single tissue compartment model.4
Results: For the HRRT, good recoveries of the spheres were obtained for both standard OP-OSEM (0.84 to 0.97) and PVC OP-OSEM (0.91 to 0.98) for the HRRT. In addition, for the HR+, good recoveries were found for the PVC-OSEM reconstruction (0.84 to 0.94), which corresponded well with results of the standard HRRT OP-OSEM reconstructions. In contrast, much lower recoveries were found for standard NAW-OSEM 5 mm (0.42 to 0.86). In addition, for clinical data, good correspondence (slope: 1.00±0.08; R2: 0.95±0.01) was found between HR+ PVC-OSEM and HRRT OP-OSEM derived VT values, see also Figure 1.
HRRT OP-OSEM versus HR+ PVC-OSEM based regional BFM VT values for 15 ROIs per subject (n = 5).
Conclusions: The present study showed that HR+ image resolution using PVC-OSEM was comparable to the resolution of the HRRT scanner. Outcome of tracer kinetic analysis of HR+ studies reconstructed with PVC-OSEM correlated well with outcome of HRRT studies, indicating that reconstruction based partial volume correction yields quantitatively accurate images.
Acknowledgments: Financial support provided by the Netherlands Organisation for Scientific Research (NWO, VIDI Grant 016.066.309.
130. Evaluation of a 3D iterative reconstruction algorithm that allows for negative image values in high resolution pet studies
F.H.P. van Velden1, C. Comtat2, J. Nuyts3, A. Reilhac4, A.A. Lammertsma1 and R. Boellaard1
1Department of Nuclear Medicine & PET Research, VU University Medical Center, Amsterdam, The Netherlands; 2Service Hospitalier Frédéric Joliot (S.H.F.J.), Commissariat à l'Énergie Atomique, Orsay, France; 3Katholieke Universiteit Leuven, U.Z. Gasthuisberg, Leuven, Belgium; 4CERMEP, Bron, France
Background and aims: The High Resolution Research Tomograph (HRRT, CTI/Siemens) is a dedicated human brain positron emission tomography (PET) scanner. At present, all available 3D iterative reconstruction algorithms show bias in low count regions of short duration frames (10 to 60 secs), hampering quantitative accuracy especially when using reference tissue models.1 In a previous simulation study,2 a maximum-likelihood iterative reconstruction method that allows for negative image values (NEG-ML)3 showed promising results with respect to bias reduction in low count regions. The goal of the present study was to evaluate quantitative accuracy of 3D ordinary-Poisson (using prompts and randoms) NEG-ML (NEG-OP) and its impact on kinetic analysis of HRRT data.
Methods: A homogeneous phantom, a 3D anthropomorphic human brain (Hoffman) phantom and dynamic [11C]flumazenil human brain studies (n = 5) were reconstructed using NEG-OP (15 iterations, 16 subsets). For comparison, reconstructions were also performed using standard 3D OP ordered subsets expectation maximization (OP-OSEM, 16 iterations, 16 subsets), recommended 3D ordered subsets weighted least squares (OSWLS, 7 iterations, 16 subsets)1 and analytical 3D filtered-backprojection (3D-FBP). Bias was measured in low count frames, using a high count frame as reference. Human brain studies were analyzed using receptor parametric mapping, a basis function implementation of the simplified reference tissue model, with pons as reference tissue, providing binding potential (BPND) images. In addition, volume of distribution (VT) and delivery (K1) images were obtained using a basis function implementation of the standard single tissue plasma input compartment model.
Results: Using NEG-OP, bias in the homogeneous phantom was reduced to within 3% for short duration frames (OP-OSEM: <4%, OSWLS: <8%, 3D-FBP: <1%), having the lowest coefficient of variation within the phantom (<150%, OP-OSEM: <410%, OSWLS: <190%, 3D-FBP: <1100%). For short duration frames of the Hoffman phantom, however, bias was still observed in both grey (<9%, OP-OSEM: <7%, OSWLS: <9%, 3D-FBP: <1%) and white (<15%, OP-OSEM: <21%, OSWLS: <28%, 3D-FBP: <1%) matter regions. Visually, NEG-OP showed less artifacts in VT, BPND and especially K1 images than the other methods. However, for both VT and BPND, linear regression with fixed intercept against 3D-FBP showed slightly poorer slopes (0.87 and 1.20, respectively) than the other iterative methods (OP-OSEM: 0.92 and 1.12, OSWLS: 0.93 and 1.02). As 3D-FBP images were very noisy, no regression for K1 images was possible.
Conclusions: Although NEG-OP is not free of bias, it does show bias reduction combined with low noise levels, resulting in improved K1 images. As 3D-FBP reconstructions are very noisy, linear regression against 3D-FBP might not provide a fair comparison. Further studies are needed to fully assess the potential of the NEG-OP algorithm for HRRT reconstructions.
Acknowledgments: Financial supported provided by the Netherlands Organization for Scientific Research (NWO, VIDI Grant 016.066.309).
844. Effect of image reconstruction algorithms on binding potential calculations in [18F]-fallypride PET
J. Dunn1, P. Marsden1, M. O'Doherty1, M. Cleij1 and L. Reed2
1PET Imaging Centre, Kings College London; 2Section of Addiction Neurobiology, Institute of Psychiatry, Kings College London, London, UK
Objectives: The PET radioligand [18F]-fallypride binds to D2/D3 receptors with high affinity in both striatal and extrastriatal regions.1 We investigated the effect of PET reconstruction algorithms on the quantification of binding potential (BP) in selected regions.
Methods: Six healthy male volunteers were injected with 250MBq of [18F]-fallypride and had a brain scan for 90 mins on a GE Discovery STE PET camera with a 3D acquisition. CT attenuation corrected dynamic images were reconstructed after 2D Fourier rebinning (FORE) with filtered back projection (FBP) and iterative (FORE-ITER) reconstruction algorithms. Also, VUE Point, a fully 3D iterative algorithm (3D-ITER), was used to reconstruct both attenuation corrected (AC) and non-attenuation corrected (NAC) dynamic PET images. The frames in each NAC 3D-ITER image were realigned using the rigid-body realignment algorithm in SPM5 (www.fil.ion.ucl.ac.uk/spm) and these transforms were applied to the FBP, FORE-ITER and AC 3D-ITER images of each subject. Time activity curves from these images were extracted from regions defining the cerebellum, amygdala, caudate, putamen, thalamus, mid temporal lobes and inferior temporal lobes, all bilaterally, from the Automated Anatomical Atlas.2 The cerebellum was used a reference region to calculate the binding potential using the simplified reference tissue models of Lammertsma and Hume.3
Results: When corrected for multiple comparisons repeated measures ANOVA revealed a significant effect of reconstruction method on BP in bilateral amygdala, right putamen and all temporal regions (see Table). Post-hoc analysis revealed significant differences in BP (denoted * in the Table) between the FORE-ITER and 3D-ITER images in bilateral amygdala, right putamen, bilateral mid temporal lobes and right inferior temporal lobe.
Descriptive statistics and comparisons of BP
Region name
Mean (s.d.) BP across subjects
Mean fractional difference in BP (%)
FBP
FORE-ITER
3D-ITER
FBP v FORE-ITER
FORE-ITER v 3D ITER
FBP v 3D-ITER
Amygdala L
2.91 (0.43)
3.01 (0.44)
2.73 (0.43)
−3.3
9.9*
6.7
Amygdala R
3.10 (0.47)
3.18 (0.47)
2.93 (0.43)
−2.5
8.4*
5.8
Putamen R
14.10 (1.59)
13.94 (0.67)
14.86 (0.97)
1.0
−6.3*
−5.3
Mid Temp. Lobe L
0.67 (0.24)
0.62 (0.23)
0.67 (0.22)
8.1
−9.3*
−1.3
Mid Temp. Lobe R
0.63 (0.20)
0.59 (0.20)
0.64 (0.19)
6.8
−9.0*
−2.3
Inf. Temp. Lobe L
0.90 (0.28)
0.84 (0.26)
0.88 (0.26)
6.7
−4.81
1.9
Inf. Temp. Lobe R
0.84 (0.26)
0.79 (0.25)
0.83 (0.25)
6.1
−4.7*
1.4
Conclusions: The spatially variant convergence of iterative reconstruction algorithms is known to affect quantified PET especially in regions of low activity.4 In most of the regions investigated BP was affected by reconstruction method, with the largest pairwise fractional differences between FORE-ITER and 3D-ITER, approaching 10% in the amygdala and mid temporal lobes. In conclusion choice of reconstruction algorithm is important especially between choices of iterative algorithms.
311. Radioligand discovery and development through bio-mathematical modelling
Q. Guo1, M. Brady1, C. Salinas2 and R. Gunn1,2
1Department of Engineering Science, University of Oxford, Oxford; 2Clinical Imaging Centre, GSK, London, UK
Objectives: The development of PET radioligands for novel neuroreceptor, transporter and enzyme targets is a complex process. Traditional radioligand screening methods focus on lipophilicity and affinity, but these only partially identify the characteristics of a successful radioligand. We have developed a bio-mathematical modeling approach which also accounts for non-specific binding and kinetics, aiming to predict the in vivo performance of candidate radioligands from in silico/in vitro data. The performance of our method is evaluated on a dataset including in silico/in vitro and in vivo data for each compound (n = 28).
Methods: The approach presented here uses a standard input function and a one tissue compartment model to approximate the in vivo behavior of ligands in both target and reference regions with a parsimonious parameter set (influx rate constant K1, efflux rate constant k2 and binding potential BPND) predicted from in silico/in vitro data. K1 prediction is based on the Renkin-Crone model (K1 = F·(1−e−PS/F)) with perfusion F, capillary surface area S, and incorporates a model to predict permeability P from lipophilicity and molecular volume. k2 is predicted by k2 = Vaq_T ·K1·fND/Vaq_P·fP, under the assumption of passive diffusion, from K1, apparent aqueous volumes in plasma Vaq_P and tissue Vaq_T, free fractions in plasma fP and tissue fND measured using equilibrium dialysis assays. The model to predict BPND (BPND = fND·Bmax/KD) uses target density Bmax and affinity KD measurements derived from in vitro homogenate binding assays and fND. The in vivo performance of a ligand is estimated using the coefficient of variation of BPND (%COV[BPND]) metric derived from Monte Carlo simulations. 28 compounds for 10 targets were evaluated using our method to predict their in vivo performance and subsequently validated against in vivo PET data in the Landrace Pig.
Results: The predicted K1, k2 and BPND values from in silico/in vitro data were consistent with the in vivo estimates (Pearson's rK1 = 0.54, pK1 = 0.001, rk2 = 0.73, pk2<0.0001, and rBPND = 0.82, pBPND<0.0001). The prediction showed that widely accepted ‘good’ ligands such as 11C-Flumazenil had small %COV[BPND] values whereas ‘poor’ imaging probes were identified with a higher value such as 11C(R)-PK11195. The model's ranking of the candidates within a specific target was generally consistent with historical decisions made on the in vivo PET data.
Conclusion: Bio-mathematical modelling can aid the radioligand discovery and development process and efficiently leverage large compound databases.
