2019 Volume 37 Issue 1 Pages 35-45
This paper describes the noise removal effect of low count PET images by using various convolutional neural networks (CNN). As the architecture of CNN, we used Remez 's DenoiseNet (DN) and U-Net with residual learning (UR-Net). In the results, streak artifacts were appeared in the coronal plane by using DN. On the other hand, the artifact was reduced by DN-Nch which removes noise by using adjacent N slices as N channel images. Furthermore, UR-Net and UR-Net×2 which is stacked in two layers improved peak signal to noise ratio (PSNR) compared to DN. In addition, it was indicated that a blind noise removal which removes noise with unknown noise level could be possible by training image dataset which is different thinning rates.
