Penalized PET reconstruction using CNN prior

Inspired by great performance of convolutional neural network (CNN), we propose an iterative positron emission tomography (PET) reconstruction using a CNN prior. We used the denoising CNN (DnCNN) method and trained the network using regular dose images as groundtruth and low dose images as input. Poisson thinning method is used for generating the low dose data by downsampling counts. Due to the DnCNN is trained at a certain noise level, the noise level change in each iteration is one of major problems. To address this issue, we propose a local linear fitting (LLF) function incorporated with DnCNN to improve the image quality by preventing unwanted bias. By using LLF function, we demonstrate that the proposed method is robust to noise level changes in iterations. In bias and variance studies in simulations, the proposed method outperforms the conventional iterative methods. We confirm that the proposed method improves the reconstructed image both quantitatively and qualitatively.