Accuracy and variability of quantitative measurements using PET with time‐of‐flight information and resolution modelling Ian S. Armstrong 1,2 , Heather A. Williams 1 , Julian C. Matthews 2 1 Nuclear Medicine Centre, Central Manchester University Hospitals, Manchester, United Kingdom 2 School of Cancer and Enabling Sciences, MAHSC, University of Manchester, Manchester, United Kingdom Introduction PET images reconstructed using time‐of‐flight (TOF) information have been shown to give greater lesion detection [1] and improved signal‐to‐noise ratio [2] compared with conventional OSEM reconstruction. Similar findings have been shown for PET images reconstructed using Resolution Modelling (RM) [3], and those reconstructed using both RM and TOF information [4]. Of these studies, only one [3] has used replicate images to assess the quantitative measurements over a range of reconstruction iterations. However, the work only studied the impact of RM and while it did report on the quantitative accuracy of measurements, it did not address variability in these measures – an important factor to consider clinically when monitoring small changes in uptake between patient studies. The aim of this study is to assess the benefits of combined RM and TOF information on the accuracy and variability of image‐derived quantitative measurements, using replicate images. Method A 60 minute listmode PET acquisition of the NEMA image quality phantom was performed on a Siemens Biograph mCT∙S. All six spheres in the phantom were filled with 18 F at activity concentrations (AC) of 41.7 kBq/ml followed by 21.7 kBq/ml, with a background AC of 5.3 kBq/ml in both cases resulting in 8:1 and 4:1 contrast ratios respectively. The phantom was positioned so that the plane through the centre of the six spheres was at the centre of PET axial field of view and the centre of the lung insert was at the centre of the transaxial field of view. ECG gating, using an ECG simulator, was incorporated into the acquisition to enable replicate images to be created from the data with consistent count statistics in each image. 10 replicate sinograms were produced with ~6.0×10 7 true coincidence events in each sinogram. Sinograms were reconstructed using 3D‐OSEM (OSEM), 3D‐OSEM with RM (Siemens HD) [5] and 3D‐OSEM with RM and TOF information (Siemens UHD). Reconstruction parameters were 1 to 12 iterations, 24 subsets for OSEM and HD, 21 subsets for UHD and a 256×256 image matrix giving voxel dimensions of 3.2×3.2×2.0 mm 3 . All images included CT‐ based attenuation correction and scatter correction. Images were reconstructed without a post‐filter, with two additional image sets created post‐reconstruction by smoothing the images with 3D Gaussian filters of 3 mm and 5 mm FWHM. In summary: 3 algorithms × 3 post filters × 12 iterations × 10 replicates per iteration = 1080 reconstructed datasets produced and analysed for each experiment. Circular regions of interest (ROIs), with diameters equal to the physical diameter of the spheres, were placed on the transaxial image slice that intersected the centre of the spheres. For each sphere, the AC recovery was measured for both the mean of voxel values and the maximum voxel value within the ROI. The mean bias and standard deviation of the measured AC was calculated across the ten replicate images and expressed as a percentage of the actual AC of the spheres. This metric can be used to assess the mean square error of the measurements. Results Plots highlighting the key features of the dataset are shown in Figures 1 and 2. A negative bias was seen for mean AC recovery measurements of all sphere sizes for both 4:1 (see Figure 1) and 8:1 contrast ratios. UHD gave better AC recovery than OSEM over all iterations and HD was better than OSEM apart from at 1 iteration, where OSEM gave better recovery. For both contrast ratios, UHD gave better mean AC recovery for the 10 mm sphere than HD or OSEM to the degree that, up to 2 iterations, recovery for UHD with a 3 mm filter was better than HD without a filter. For larger spheres, mean AC recovery was only better with UHD compared with HD for low iterations due to faster convergence. At increasing numbers of iterations, the two algorithms were either equivalent or HD gave better recovery – this is seen in Figure 1. This suggests that for low numbers of iterations, as used for typical clinical images, UHD will offer superior AC recovery than both HD and OSEM. A positive bias was seen for nearly all measures of maximum AC recovery. For the 10 mm sphere, a greater bias was observed with UHD compared with HD for all iterations in the 4:1 phantom and up to 5 iterations in the 8:1 phantom. OSEM gave the lowest positive bias in both 4:1 and 8:1 phantoms. This was reversed for the 37 mm sphere, where OSEM gave the greater positive bias compared with HD and UHD, which gave a small positive bias of comparable magnitude. Trends for the other spheres fell between these two extremes.