Low Dose Brain PET Imaging using Denoising Diffusion Probabilistic Models

Moderating the injected dose to reduce the potential health hazards has been always a desirable goal that can reduce acquisition time and can have a positive impact on the patient’s comfort and on the scanner’s throughput. Reducing the scanning time and injected dose have a direct impact on image quality and can lead to hampering of signal-to-noise ratio, increase quantitative bias and reduce the diagnostic value of PET imaging. A novel deep learning image-to-image translation method called Denoising Diffusion Probabilistic Models (DDPMs) was utilized in this work to improve the quality of low-dose (LD) PET images. A dataset including 41 ¹⁸F-FDG PET images belonging to patients with cognitive symptoms of the possible neurodegenerative disease was reconstructed to generate the full dose (FD) images and a subset of PET data containing 5% of the total events was considered as LD images. Quantitative analysis includes established metrics such as the peak signal-to-noise ratio (PSNR), structural similarity index metric (SSIM), and root means square error (RMSE) were calculated. The result shows that the model can increase the SSIM and RMSE and improve overall image quality. The suggested DDPMs denoising model enables us to reduce the radiation dose to patients to only 5% and increase its comfort in the case of fast scanning.