Motivation: Synthesizing full-dose PET images from MRI data holds significant clinical potential by reducing patients' radiation exposure during imaging procedures. The use of generative-AI in epilepsy imaging remains underexplored. Goal(s): We compared the performance of generative and non-generative deep learning models in MRI-PET image translation tasks through clinically focused evaluations. Approach: We adapted existing 2D cross-modal image translation models including CNN-based Transformer-UNet and Score-based Generative Models trained on axial slices collected from PET-MR. Results: We demonstrate the ability to synthesize full-dose FDG-PET images from MRI inputs alone using deep learning diffusion models and demonstrate further improved performance with ultralow-dose (1%) FDG-PET as an input. Impact: This study suggests the possibility to use generative AI approaches to massively reduce dose levels for FDG PET brain studies. Further work will leverage 3D patch-based approaches can improve the performance and slice consistencies.
Wu et al. (Tue,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: