Ultra-low count total-body PET image denoising based on the deep learning method
Objective To improve the reconstructed image quality of low-count positron emission tomography(PET)imaging based on deep learning method and explore the generalization performance of the proposed method on different noise levels.Methods Using the dataset from the MICCAI 2022 UDPET Challenge,the hierarchical vector quantized variational autoencoder(HVQ-VAE)method was proposed to denoise low-count PET images with different dose reduction factors(DRFs).The denoising efficacy was quantitatively evaluated via metrics such as normalized root mean square error,structural similarity,and peak signal-to-noise ratio,as well as through visual assessments,against the Gaussian filter as baseline mothod.Results When the DRF of low-count PET images was 20,the overall image quality was improved by 13%after Gaussian filtering,and 20%after denoising by HVQ-VAE.At a DRF of 50,the proposed approach outperformed the Gaussian filter,delivering a 24%quality improvement compared to its 11%.At the DRF of 100,the HVQ-VAE method marked 36%improvement in overall image quality,as opposed to the 12%achieved with the Gaussian filter.Conclusion The HVQ-VAE method,as part of our proposed technique,has demonstrated a marked denoising effect on total-body ultra-low-count PET images across diverse noise levels.This research opens up novel avenues for reducing radiation exposure risks while ensuring maintenance of image fidelity.
positron emission tomographyimage denoisinglow-count PET imagetotal-body PET imageHVQ-VAE
万方数据
维普
