Low-count PET image denoising based on VQVAE: evaluation on total-body datasets

Abstract

Enhancing the quality of low-count positron emission tomography (PET) images is clinically significant. This study aims to produce high-quality results from low-count total-body PET images using generative deep learning methods for image denoising. This work proposed a new generative model, which integrates attention mechanism and dilated convolution into the vector quantized variational autoencoder (VQVAE) framework, which was named LAHVQVAE. Utilizing 320 total-body patient data pairs from the 2022 MICCAI Ultra-Low Dose PET Imaging Challenge (210 for training, 30 for validation, and 80 for testing), The low-count total-body PET images with a dose reduction factor (DRF) of 20 and the standard-count total-body PET images serve as the input and label of the network, respectively. Compared with Gaussian filter, VQVAE, and image space shuffle U-Net (ISSUNET), the LAHVQVAE achieved the lowest average MAE of 0.0123, the highest average PSNR of 62.51 and the highest average SSIM of 0.9975. Wilcoxon signed-rank test results indicate that the proposed model significantly outperforms other methods in terms of PSNR (p<0.001). Distribution analysis shows that the output images of proposed method are more concentrated in low MAE, high PSNR and high SSIM compared other methods. The LAHVQVAE provides effective image denoising results, outperforming the Gaussian filter, VQVAE, and ISSUNET in qualitative, quantitative, and distribution analyses.