Feature Loss After Denoising of SPECT Projection Data using a U-Net

It has been long known that denoising of Single Photon Emission Computed Tomography (SPECT) projection data from multi channel collimators harbors the danger that features are filtered away as distinction between noise and signal in the 2D projection image does not consider the tomographic information (see Vija, Yahil et al.). We investigate if an approach using Neural Networks can overcome the past issues, by training a U-Net on Monte Carlo Simulations of the XCAT phantom. Although error metrics indicate improvement of signal statistics, we observe a loss of image features after the reconstruction. We show that when using OSEM reconstruction without denoising in projection domain, these features can be preserved.