Low-count whole-body PET/MRI restoration: an evaluation of dose reduction spectrum and five state-of-the-art artificial intelligence models

Anatomical-guided attention enhances unsupervised PET image denoising performance

“…We evaluated the proposed method using clinical brain PET data with various radioactive tracers, 18 F-florbetapir (AV-45), 11 C-Pittsburgh compound-B (PIB), 18 F-fluoro-2-deoxy-Dglucose (FDG), and 15 O-CO2 acquired from different PET scanners. Note that our ultimate goal is not the task of image restoration using self-supervised learning.…”

“…For the evaluation using 18 F-AV-45, 11 C-PIB, and 18 F-FDG, we used cognitively normal subjects from the Open Access Series of Imaging Studies (OASIS-3), a public dataset containing multi-modal information from 42 to 95 years of age [36]. Table 1 summarizes the evaluation datasets.…”

“…Attenuation correction was performed by using a μmap from Dixon-based tissue segmented MR scan. Participants using 11 C-PIB and 18 F-FDG received 6 mCi to 20 mCi and approximately 5 mCi, respectively, followed by a dynamic 60 min emission scan. The 11 C-PIB scan was performed using Siemens Biograph 40 and reconstructed using a discrete inverse Fourier transform (DIFT) after a spiral CT scan for attenuation correction.…”

“…We used only 18 F-AV-45 data from 24 subjects for selfsupervised pre-training. DIP-based PET image denoising of the main task was evaluated using 18 F-AV-45, 11 C-PIB, and 18 F-FDG data from every 10 subjects, but not from the pre-training dataset. As the ultra-low dose 15 O-CO2 data were generated by downsampling the full list-mode data, we generated independent 16 noisy samples and evaluated their denoising performance.…”

“…Onishi, F. Hashimoto, and K. Ote are with the Central Research Laboratory, Hamamatsu Photonics K.K., Hamamatsu, 434-8601, Japan developed rapidly over the past several years [6]. The general approach is to map low-quality PET images to high-quality PET images through supervised training using a convolutional neural network (CNN) [7]- [9] or a transformer model [10], [11].…”