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

Accurate attenuation correction (AC) is critical in quantitative brain PET imaging. Conventional CT-based AC methods increase radiation exposure and workflow complexity, and thereby create barriers to the widespread adoption of dedicated brain PET scanners. Deep-learning-based AC, although promising, frequently lacks generalizability to tracers that are not included in the training dataset. This domain-shift problem implies that, in practice, separate models might be needed for each tracer or clinical indication, which is impractical, especially for newly developed tracers and rare diseases where training data are scarce. This study aimed to develop a tracer-agnostic deep-learning approach for a CT-free AC framework.

The proposed method employed a denoising diffusion probabilistic model to generate pseudo-transmission CT images from non-AC PET images. Two simple strategies were introduced to enhance cross-tracer generalizability: a visualtransformation module that synthesizes diverse tracer-dependent contrasts, and slice-positional embeddings to maintain anatomical continuity. The model was trained on 181 [¹⁸F]FDG datasets and comprehensively evaluated in 14 multi-tracer datasets acquired on a dedicated brain PET scanner.

The proposed method outperformed the emission-segmented AC and conventional U-Net models and achieved superior generation accuracy and regional bias within 10% across all tracers. Robust generalizability was demonstrated in tracers with diverse uptake patterns and acquisition protocols.

This CT-free AC approach achieves quantitative accuracy comparable to that of CT-based correction while eliminating additional scans, reducing radiation exposure, and simplifying the workflow. This tracer-independent design supports the early evaluation of novel tracers and their broad clinical adoption in brain-PET imaging.