MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Positron Emission Tomography Scatter Image Reconstruction with CNN Machine Learning

G Fontaine*, S Pistorius, University of Manitoba, Winnipeg, MB

Presentations

(Sunday, 7/12/2020)   [Eastern Time (GMT-4)]

Room: AAPM ePoster Library

Purpose: To implement a machine learning convolutional neural network to directly reconstruct positron emission tomography (PET) images from raw sinogram data. The network will be trained to incorporate scattered coincidences within the reconstruction.

Methods: PET activity uptake images and their corresponding sinograms were simulated to train and test a convolutional neural network (CNN). The activity uptake images were generated with XCAT software from Duke University. 180,000 unique two-dimensional images were created for training purposes. Sinograms were simulated with analytical and Monte Carlo simulations, and analytical simulations were created from the inverse radon transform with the addition of Poisson noise. Monte Carlo simulations were generated with Geant4 Application for Tomographic Emission (GATE). From the GATE simulations, the sinograms were sorted into energy-dependent sinogram bins with 50 keV widths. The CNN was trained using transfer learning by initially training the network with the analytically simulated sinograms. Following this, a portion of the CNN was frozen, and the remaining layers were re-trained with the energy-dependent sinograms to incorporate scattered coincidences.

Results: The convolutional neural network was trained with the analytical sinogram simulations to reconstruct PET images. The network was validated during training with a validation set to monitor under- or over-fitting. The CNN was tested with an isolated set of sinogram-image pairs, and the reconstructed images were compared to images reconstructed using filtered-back projection (FBP). The mean structural similarity to the original images was found to be 0.82(9) and 0.69(7) for CNN and FBP reconstructions, respectively.

Conclusion: A convolutional neural network may be trained to incorporate scattered data during PET image reconstruction. This method overcomes some of the issues associated with analytical scatter reconstruction and may improve image quality, decrease computational time, and maximize the amount of useful data for a given dose.

Keywords

PET, Reconstruction, Scatter

Taxonomy

IM- PET : Image Reconstruction

Contact Email