DG-DiT: Dual-Branch Gating Diffusion Transformer for Multi-Tracer and Multi-Scanner Brain PET Image Denoising.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiochemistry

Background:

• Diffusion probabilistic models (DPMs) show promise for positron emission tomography (PET) denoising but face challenges in efficiency, accuracy, and generalizability.
• Existing DPMs often require many iterations, exhibit limited quantitative accuracy, and struggle with data variability from multiple scanners and tracers.
• High fidelity to ground truth is crucial for PET denoising, necessitating improved model performance.

Purpose of the Study:

• To develop an efficient and accurate deep learning model for multi-tracer and multi-scanner PET denoising.
• To address limitations of traditional DPMs in quantitative accuracy and generalizability across different imaging conditions.
• To improve the modeling of complex data distributions in PET imaging.

Main Methods:

• Proposed a dual-branch gating diffusion transformer (DG-DiT) network integrating a diffusion transformer (DiT) and an image restoration transformer (IRT).
• The DiT backbone learns priors from a compact latent space for efficient few-step diffusion.
• A dual-branch gating mechanism is employed in both DiT and IRT to fuse multi-input information effectively.

Main Results:

• The DG-DiT model achieved superior quantitative accuracy across all tested scanners and tracers, showing up to 0.2 dB PSNR improvement over state-of-the-art models.
• Contrast-to-noise ratio evaluations demonstrated the model's ability to recover contrast in small brain regions while reducing noise.
• Extensive experiments on multi-tracer and multi-scanner datasets validated the model's consistent denoising performance.

Conclusions:

• The proposed DG-DiT network offers a significant advancement in PET image denoising, particularly for multi-tracer and multi-scanner scenarios.
• DG-DiT provides enhanced quantitative accuracy and superior noise reduction capabilities compared to existing deep learning methods.
• The model's architecture facilitates efficient learning and improved generalizability for diverse PET imaging data.