Authors:
Gabrielle Fontaine
1
;
Peter Lindstrom
1
and
Stephen Pistorius
1
;
2
Affiliations:
1
Department of Physics and Astronomy, University of Manitoba, 30A Sifton Road, Winnipeg, MB R3T 2N2, Canada
;
2
CancerCare Manitoba Research Institute, 675 McDermot Ave, Winnipeg, MB R3E 0V9, Canada
Keyword(s):
Positron Emission Tomography, Medical Imaging, Neural Network, Deep Learning, Scattered Coincidences, Direct Reconstruction.
Abstract:
PET image reconstruction largely relies on pre-reconstruction data correction, which may add noise and remove information. This loss is particularly notable when correcting for scattered coincidences, which are useful for image reconstruction, though algorithmic scatter reconstructions require a detector energy resolution that exceeds the current state-of-the-art. Preliminary research has demonstrated the feasibility of using convolutional neural networks (CNNs) to reconstruct images directly from sinogram data. We have extended this approach to reconstruct images from data containing scattered coincidences. Monte Carlo simulations were performed to simulate PET data from digital phantoms. Data were modeled using 15% FWHM energy resolution detectors. Energy-dependent sinograms (EDSs), containing true and scattered coincidences, were constructed from the data. After data augmentation, 210,000 sinograms were obtained. A CNN was trained on the EDS-activity pairs for image reconstruction
. A second network was trained on sinograms containing only photopeak coincidences. Images were also reconstructed using FBP, and MLEM approaches. The EDS trained network outperformed the photopeak trained network, with a higher mean structural similarity index (0.69 ± .05 vs. 0.63 ± .05) and lower average mean square error (0.16 ± .04 vs. 0.20 ± .04). Our work demonstrates that CNNs have the potential to extract useful information from scattered coincidences, even for data containing significant energy uncertainties.
(More)