Dose-aware diffusion model for 3D PET image denoising: Multi-institutional validation with reader study and real...

Area of Science:

• Medical Imaging
• Radiology
• Artificial Intelligence in Medicine

Background:

• Low-count/low-dose Positron Emission Tomography (PET) imaging requires reduced scan times and radiation dose while maintaining image quality, especially for lower-performance scanners.
• Existing deep learning (DL) denoising models for PET often compromise image quality, lack generalizability across different noise levels, acquisition protocols, and patient populations.
• While diffusion models show promise for high-quality sample generation in medical imaging, current applications in low-dose PET struggle with 3D reconstruction consistency, generalization, and accurate detail rendering.

Purpose of the Study:

• To develop and evaluate DDPET-3D, a dose-aware diffusion model designed for 3D-consistent low-dose PET imaging.
• To address limitations of existing diffusion models in generating consistent 3D reconstructions, generalizing across noise levels, and avoiding distorted details in low-dose PET.
• To validate the model's performance using a large, multi-center dataset and assess its potential for clinical application.

Main Methods:

• Development of DDPET-3D, a dose-aware diffusion model utilizing a 2.5D conditioning backbone for 3D-consistent reconstruction.
• Extensive evaluation using 9783 18F-FDG PET studies from 4 global medical centers, covering low-dose levels from 1% to 50% and diverse scanners/protocols.
• Cross-center and cross-scanner validation to assess model generalizability, supplemented by reader studies and Monte Carlo simulations for quantitative accuracy.

Main Results:

• DDPET-3D demonstrated strong generalizability across different low-dose levels, scanners, and clinical protocols, validated across multiple centers.
• Nuclear medicine physicians rated denoised images from DDPET-3D as comparable or superior to full-dose images and existing DL baselines in qualitative assessments.
• Quantitative evaluation using Monte Carlo simulations and a lesion segmentation network indicated accurate lesion-level quantitative accuracy, supporting the potential for low-dose PET with maintained image quality.

Conclusions:

• DDPET-3D effectively addresses challenges in 3D low-dose PET imaging, offering consistent and high-quality reconstructions.
• The model shows significant potential for clinical translation, enabling reduced radiation dose in PET scans without compromising diagnostic image quality.
• Public availability of code and models facilitates further research and application of advanced diffusion models in medical imaging.