Tracer-agnostic diffusion model-based CT-free attenuation correction for brain PET: Comprehensive evaluation across...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Nuclear Medicine

Background:

• Accurate attenuation correction (AC) is crucial for quantitative brain PET imaging.
• Conventional CT-based AC increases radiation exposure and workflow complexity, hindering PET scanner adoption.
• Current deep-learning AC methods struggle with generalizability across different tracers, necessitating tracer-specific models.

Purpose of the Study:

• To develop a tracer-agnostic deep-learning framework for CT-free attenuation correction (AC) in brain PET.
• To overcome the limitations of existing AC methods regarding generalizability and workflow efficiency.
• To enable wider adoption of dedicated brain PET scanners and facilitate the evaluation of novel tracers.

Main Methods:

• A denoising diffusion probabilistic model was used to generate pseudo-transmission CT images from non-AC PET images.
• Strategies including a visual-transformation module and slice-positional embeddings were implemented to enhance cross-tracer generalizability.
• The model was trained on [18F]FDG datasets and validated on 14 multi-tracer datasets from a dedicated brain PET scanner.

Main Results:

• The proposed CT-free AC method demonstrated superior generation accuracy and regional bias (<10%) compared to emission-segmented AC and U-Net models.
• Robust generalizability was confirmed across various tracers with diverse uptake patterns and acquisition protocols.
• Quantitative accuracy comparable to CT-based correction was achieved without additional scans.

Conclusions:

• The developed CT-free AC approach enhances quantitative accuracy while simplifying workflows and reducing radiation exposure.
• This tracer-independent deep-learning method supports the efficient evaluation of novel PET tracers.
• The framework promotes broader clinical adoption of brain PET imaging by removing barriers associated with conventional AC.