Evaluating AI-aided approaches for 18F-FDG PET quantification: Indirect synthetic MR-based versus direct partial volume correction

Compatible deployment of AI-aided methods for PET quantification is an important prerequisite to maximizing their benefits. We focus on partial volume correction (PVC), which can substantially improve the precision of radiotracer uptake quantification in brain PET for intricate and atrophic regions. Conventional post-reconstruction PVC requires anatomical MR images that are often unavailable or of inadequate quality. We address this limitation by systematically evaluating two AI-aided strategies: (1) indirect PVC, which uses synthesized MR images for anatomical guidance, and (2) direct PVC, which predicts corrected PET images without anatomical processing. Multiple AI architectures were assessed under both strategies, using paired 18 F-FDG PET + CT + MR datasets from multiple scanner sites. Indirect PVC consistently outperformed direct approaches across all tested architectures with the Denoising Diffusion Probabilistic Model yielding the best overall performance while preserving compatibility with standard PET processing pipelines. Both AI-aided approaches increased the utility of standalone 18 F-FDG PET in clinical and research applications without requiring high-resolution MR images. Indirect PVC showed advantages in transparency and performance for quantification in smaller anatomical regions, whereas direct PVC may be more suitable for rapid assessment in the larger brain regions.