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

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Partial volume correction (PVC) is crucial for accurate brain PET quantification, particularly in complex or atrophic regions.
• Conventional PVC methods often rely on anatomical MR images, which may be unavailable or of poor quality.
• AI-driven strategies offer potential solutions to overcome limitations of traditional PVC.

Purpose of the Study:

• To evaluate the effectiveness of two AI-aided PVC strategies: indirect PVC (using synthesized MR images) and direct PVC (predicting corrected PET images).
• To compare the performance of various AI architectures under both strategies for 18F-FDG PET quantification.
• To assess the utility of AI-aided PVC for standalone 18F-FDG PET applications.

Main Methods:

• Systematic evaluation of indirect and direct AI-aided PVC strategies.
• Assessment of multiple AI architectures using paired 18F-FDG PET/CT/MR datasets from multi-site scanners.
• Comparison of performance metrics for quantification accuracy and applicability.

Main Results:

• Indirect PVC consistently outperformed direct PVC across all tested AI architectures.
• The Denoising Diffusion Probabilistic Model demonstrated the best overall performance within the indirect PVC strategy.
• Both AI-aided approaches improved the utility of standalone 18F-FDG PET, reducing reliance on high-quality MR images.
• Indirect PVC showed superior transparency and performance for small anatomical regions, while direct PVC was suitable for rapid assessment of larger regions.

Conclusions:

• AI-aided PVC, particularly indirect PVC using synthesized MR images, significantly enhances brain PET quantification accuracy.
• The Denoising Diffusion Probabilistic Model is a promising AI architecture for indirect PVC, compatible with standard PET processing.
• AI-driven PVC methods expand the clinical and research applications of standalone 18F-FDG PET, even without high-resolution MR data.