Computational Efficient Brain PET Image Denoising by Diffusion Probabilistic Models

• Main Authors: Sanaat, A., Mazandarani, H.R., Mansouri, Z., Amini, M., Salimi, Y., Shiri, I., Zaidi, H.
• Format: Conference Proceeding
• Language: English
• Subjects: Brain modeling; Computational modeling; Noise reduction; Probabilistic logic; PSNR; Statistical analysis; Training
• ISSN: 2577-0829
• DOI: 10.1109/NSSMICRTSD49126.2023.10338186

Given the accomplishments of Denoising Diffusion Probabilistic Models (DDPMs) in a diverse range of generative artificial intelligence applications, including those related to medical imaging, their limited training and inference efficiency remain critical shortcomings. To address this issue, we propose an effective approach for generating high-quality, full-dose (FD) Positron Emission Tomography (PET) images from low-resolution, noisy low-dose (LD) PET images. Instead of initiating the denoising process by sampling from a Gaussian distribution, as in conventional DDPMs, our method commences with the output of a trained Variational Autoencoder (VAE) model, thereby enhancing computational efficiency. To assess the efficacy of our approach, we employed a dataset consisting of 75 18 F-Flutemetamol PET raw data from patients diagnosed with neurodegenerative diseases. We reconstructed FD images and corresponding LD images, which contained only 5% of the total events. Subsequently, we conducted a quantitative analysis using well-established metrics, including peak signal-to-noise ratio (PSNR), structural similarity index metric (SSIM), and root mean square error (RMSE). Our results revealed that the proposed DDPM-based denoising model can enhance SSIM and PSNR values up to 26% and 31%, respectively. This approach facilitates a reduction in the injection dose administered to patients by up to 95% while simultaneously improving quantitative and qualitative factors.