PET Image Denoising Using Unsupervised Domain Translation

Neural networks that perform domain translation have been successfully used for image restoration in settings where paired datasets are not available for network training. Here we present an unsupervised domain translation technique for PET image denoising which restores noisy real PET images by learning shared latent space information across a simulated clean image domain and a real noisy image domain. For training and validating the network in unsupervised mode, we used noisy human brain PET scans in combination with a set of noiseless simulated images based on the BrainWeb digital phantom. Using the peak signal-to-noise ratio as our evaluation metric, we show here that this unsupervised domain translation technique leads to quantitative improvements over Gaussian filtering. Unsupervised domain translation also leads to improvements in visual image quality relative to Gaussian filtering as evidenced by relative degrees of enhancement in gray matter intensities in the denoised brain PET images.