Reconstruction of Time-Of-Flight Projection Data with the STIR reconstruction framework 1 2 1 3 2 N. Efthimiou , E. C. Emond , C. Cawthorne , C. Tsoumpas and K. Thielemans 1.School of Life Sciences, Faculty of Health Sciences, University of Hull, UK 2. Inst. of Nuclear Medicine, University College of London, UK 3. Division of Biomedical Imaging, University of Leeds, UK
Objectives 1. Validate the new TOF projectors and data structures in STIR (http://stir.sf.net). 2. Present improved image quality of TOF reconstruction over non-TOF.
XCAT phantom reconstructions: Lesion Mean Value and SNR Reconstructed XCAT images of the thorax with TOF, at the 72th sub-iteration, are sharper than their non-TOF counterparts, with better defined edges.
Introduction Time-of-Flight (TOF) is a widely used method to improve PET by recording the difference in arrival time of the two annihilation photons. Image reconstruction software needs to handle this information appropriately. STIR is a well-established open source reconstruction toolkit[1]. A first work on TOF list-mode reconstruction for STIR was presented last year[2]. Here we present further updates related to TOF projection data and validate the software using analytical and Monte Carlo simulations. Materials and Methods
Reconstruction Analysis We used 2 ROIs: on the lesion located in the left lung and the whole right lung, obtaining CoV in the right lung and SNR as (lesion mean activity value)/(right lung standard deviation)
TOF kernel model: The model used to calculate the TOF kernel uses Gaussian integrated over the TOF bin [2][3].
Simulations: analytical (forward-projecting an image and including attenuation) and Monte Carlo simulations using GATE[4]. Scanner: GE Discovery PET/CT 690, time resolution of 550ps and 55 TOF bins of width 89ps. Three other time resolutions were also simulated: 300ps, 400ps and 500ps. TOF mashing was applied to reduce the number of TOF bins to 11 in some of the simulations. Poisson noise: level comparable to a real acquisition. Reconstructions: OSEM using 18 subsets, 108 sub-iterations and post-filtering was 6.4mm full width at half maximum. Validation The TOF kernel was validated with the calculation of subset sensitivity images (backprojection of 1s) using the non-TOF backprojector and TOF backprojectors for different time resolutions. Also, summing a TOF sinogram over its TOF dimension gives the same sinogram as a non-TOF forward projection (difference < 0.001%).
Figure a) - lesion mean activity vs CoV Results: lesion mean activity increases with the number of iterations, along with the CoV, as the images get noisier, with slower convergence for non-TOF. Figure b) - SNR vs sub-iteration number Original images: SNR decreases at each iteration, as images get noisier. Postfiltered images: SNR slightly increases between the 18th and the 36th sub-iterations, and then slowly decreases. At the 18th sub-iteration, SNR calculated from the filtered images is only slightly higher than from the original images, but the difference increases with iteration. Illustration of reconstruction with a Limited Field-of-View (FOV) A limited FOV (large patients, PET/MR...) can affect image quality (shadowing + wrong quantification). The following images are reconstructed and post-filtered the same way as before (72th sub-iteration) but with a smaller image size. TOF images are somewhat less affected by the truncation, as expected.
To further validate the TOF projection matrix, GATE was used to simulate point sources in different locations (time resolution: 75ps, 275 TOF bins of width 17.8ps). We compared the centre of mass of maximum probability backprojections of the ROOT data with the location of the point sources, as shown for two separate point source simulations on the figure above. Illustration of TOF Sinogram Forward Projection Figures a) and b) illustrate the forward projection to non-TOF and TOF sinograms of an emission image consisting of an oblique plane in the centre of the FOV (L40xW15xH0.8cm3).
Conclusion Significant progress has been made in STIR TOF reconstruction. Validation with analytical and Monte Carlo data show good results. Further work is in progress to support scanner data including the GE PET/MR Signa and Siemens mCT. These additions will become available in a next release. Unfortunately, TOF scatter still remains to be addressed. References [1] K. Thielemans, C. Tsoumpas, S. Mustafovic, T. Beisel, P. Aguiar, N. Dikaios, and M.W. Jacobson. STIR: Software for Tomographic Image Reconstruction Release 2. PMB, vol. 57, 867-883, 2012. [2] N. Efthimiou, K. Thielemans, C. Tsoumpas. Initial Validation of Time-Of-Flight List-Mode MLEM and OSEM Reconstruction Algorithms in STIR framework, using Monte Carlo simulated data. IEEE MIC 2016.
Cylindrical phantoms Reconstructed images of a cylindrical phantom after 18 sub-iterations show that TOF reconstruction provides sharper images with higher noise.
[3] A. Mehranian et al. Accelerated time-of-flight (TOF) PET image reconstruction using TOF bin subsetization and TOF weighting matrix pre-computation PMB, vol. 61, 1309, 2016. [4] S. Jan et al. GATE: A simulation toolkit for PET and SPECT. PMB, 49(19):4543–4561, 2004.
Acknowledgments This project was supported by the European Cooperation for Science and Technology Action TD1401: Fast Advanced Scintillation Timing (http://cern.ch/fast-cost) and by the CCP for Synergistic PET-MR Reconstruction (http://www.ccppetmr.ac.uk), EPSRC grant EP/M022587/1 and GSK funding to UCL (BIDS3000030921).
[email protected]
