Image Reconstruction Using TOF Back Projection and a Deblurring Neural Network

The increasing timing resolution of PET/CT scanners has facilitated more precise localization of annihilation events, but has yet to achieve enough certainty to produce a crisp image in one step directly from raw data. Hence simple back-projection of time-of-flight (TOF) data, while very fast, produces a blurry output called a histo-image. This article proposes a novel reconstruction pipeline consisting of a back-projection of TOF sinograms to create an intermediate histo-image and a neural network to deblur the histo-image producing a crisp high quality reconstructed image. The proposed pipeline is tested on a simulated data set of 1,083 PET images and obtains a worst case normalized mean squared error of less than 0.1% and a SSIM measure of 0.9998. These results show the promise of combining TOF back-projection and a neural network to rapidly produce high quality images.