U-Net for SPECT Image Denoising

Single Photon Emitted Computed Tomography (SPECT) is characterized by low photon counts which results in a high degree of image noise. In this work, we demonstrate the feasibility of applying deep learning-based image denoising networks to this imaging modality. We propose a four-layered U-Net for denoising of SPECT images that was trained using data obtained from Monte Carlo simulations of XCAT phantoms. Our proposed network was able to increase image quality by a factor of 14.15 dB as measured with PSNR.