Dose-aware diffusion model for 3D PET image denoising: Multi-institutional validation with reader study and real low-dose data.
Authors
Affiliations (14)
Affiliations (14)
- Department of Biomedical Engineering, Yale University, USA. Electronic address: [email protected].
- Department of Computer Science, Washington University in St. Louis, USA.
- Department of Radiology, University of California, Davis, USA.
- Department of Biomedical Engineering, Yale University, USA.
- Department of Radiology and Biomedical Imaging, Yale University, USA.
- Department of Nuclear Medicine, National Taiwan University Cancer Center, Taipei, Taiwan.
- Department of Pediatrics, Tri-Service General Hospital, Taipei, Taiwan; Department of Pediatrics, School of Medicine, College of Medicine, National Defense Medical University, Taipei, Taiwan.
- Department of Radiology, Washington University in St. Louis, USA.
- Department of Computer Science, Washington University in St. Louis, USA; Department of Electrical & Systems Engineering, Washington University in St. Louis, USA.
- Department of Nuclear Medicine, Ruijin Hospital, Shanghai Jiao Tong University School of Medicine, China.
- Department of Nuclear Medicine, University of Bern, Switzerland.
- Department of Biomedical Engineering, Rensselaer Polytechnic Institute, USA.
- Department of Radiology, University of California, Davis, USA; Department of Biomedical Engineering, University of California, Davis, USA.
- Department of Biomedical Engineering, Yale University, USA; Department of Radiology and Biomedical Imaging, Yale University, USA. Electronic address: [email protected].
Abstract
Reducing scan times, radiation dose, and enhancing image quality, especially for lower-performance scanners, are critical in low-count/low-dose PET imaging. Deep learning (DL) techniques have been investigated for PET image denoising. However, existing models have often resulted in compromised image quality when achieving low-count/low-dose PET and have limited generalizability to different image noise levels, acquisition protocols, and patient populations. Recently, diffusion models have emerged as a state-of-the-art generative model to generate high-quality samples and have demonstrated strong potential for medical imaging tasks. However, for low-dose PET imaging, existing diffusion models fail to generate consistent 3D reconstructions (i.e., adjacent slices exhibit noticeable discontinuities or "flickering" along the z-axis), struggle to generalize across varying noise levels, and often produce visually appealing but distorted details and biased tracer uptake. Here, we develop DDPET-3D, a dose-aware diffusion model for 3D low-dose PET imaging to address these challenges. In this work, "3D" denotes 3D-consistent reconstruction achieved via a 2.5D conditioning backbone, rather than a fully 3D diffusion network. Collected from 4 medical centers globally with different scanners and clinical protocols, we extensively evaluated the proposed model using a total of 9783 <sup>18</sup>F-FDG studies (1,596 patients) with low-dose/low-count levels ranging from 1% to 50%. With a cross-center, cross-scanner validation, the proposed DDPET-3D demonstrated its potential to generalize to different low-dose levels, different scanners, and different clinical protocols. As confirmed by reader studies conducted by board-certified nuclear medicine physicians, the readers rated the denoised images as comparable to or better than the full-dose images and prior DL baselines based on qualitative visual assessment. We also evaluated the lesion-level quantitative accuracy using a Monte Carlo simulation study and a lesion segmentation network. The presented results show the potential to achieve low-dose PET while maintaining image quality. Lastly, a group of real low-dose scans was also included for evaluation to demonstrate the clinical potential of DDPET-3D. Code and trained models are publicly available at https://github.com/HuidongXie/DDPET-3D.