Unpaired PET image enhancement using an improved cycle consistent adversarial network

Abstract

With the improvement of hardware technology, novel scanners such as uEXPLORER obtain significantly higher image quality than conventional scanners, but they have yet to be widespread. In addition, due to the limitations of scanning time and radiotracer dose, the PET images obtained in primary hospitals usually contains a lot of noise, which has an impact on quantitative medical analysis and diagnostic accuracy. It is of great interest to use image enhancement methods to enable primary hospitals to obtain PET images of comparable quality to the novel scanners. Obviously, in this task, we are unable to construct paired training sets, which makes common deep-learning-based methods inapplicable.To overcome this challenge, we propose to use a cycle consistent generative adversarial network (CycleGAN) that does not require paired data to achieve PET image enhancement. In addition, we add correlation coefficient loss and image prior loss to the original CycleGAN model to directly constrain the structural consistency between the generated and input images. When using the proposed model for PET image enhancement, the high-quality image dataset acquired from the novel scanners is fixed as the basic high-quality training set. Since the low-and high-quality images can be unpaired, any low-quality PET images from different institutions that require enhancement can be input into the network as the low-quality set. After training, the generator can convert low-quality images into PET images with comparable image quality to that of the novel scanners. The results show that the proposed method outperforms NLM, BM3D, and DIP. Moreover, the proposed method performs better than the supervised RED-CNN and CycleGAN when implemented on a local hospital dataset.