Intelligent medical image reconstruction: PET image denoising and super-resolution joint processing based on generating adversarial network

PET (Positron Emission Tomography) is one of the advanced imaging technologies in nuclear medicine. It is now widely used in diagnosing many diseases, determining conditions, evaluating efficacy, performing organ function research, and developing new drugs. However, due to the characteristics of nuclear medicine technology, PET is still a low-count and low- resolution imaging method. Therefore, solving the high noise and low resolution of PET images is essential. This paper proposes a multitask least square generative Adversarial Network (MultiLSGAN) for joint noise-reduction-super-resolution reconstruction of low-quality PET images. This paper proposes a Multitask Least Square Generative Adversarial Network (MultiLSGAN) for joint noise-reduction-super-resolution reconstruction of low-quality PET images. The method is targeted for noise-reduction- super-resolution imaging tasks with supervised learning. Based on the loss function of the original GAN, we introduce the VGG model to calculate the perceptual loss and add mean square error (MSE), gradient, total variance, and other loss function constraints to generate images. Finally, experiments are conducted using the data provided by Neusoft Medical, and the quantitative and qualitative results of MultiLSGAN are better compared with the advanced methods.