Dose-aware diffusion model for 3D low-dose PET denoising: A multi-institutional validation with reader study and real low-dose data

Introduction: Low-dose PET imaging is an important topic. Recently, diffusion models have demonstrated strong potential for different medical imaging tasks. However, it is challenging to extend diffusion model for 3D PET denoising problems due to the memory burden. 2D diffusion model would result in sever inconsistency between different slices. In this work, we introduce DDPET-3D, a Dose-aware Diffusion model for 3D low-dose PET imaging to address these challenges. DDPET-3D is dose-aware and can be generalized for PET image denoising of different noise-levels. We extensively evaluated the performance of DDPET-3D using a large number of ¹⁸F-FDG PET images acquired from three scanners at medical centers across three continents with low-count levels ranging from 1% to 50%. To demonstrate the clinical potential, reader studies were conducted to assess the image quality. A group of real low-dose data was also included for evaluation.

Methods: For our proposed DDPET approach, to address the 3D inconsistency problem often associated with diffusion model for 3D imaging, we utilized neighboring slices as conditional information during forward and reverse sampling steps and used multiple Gaussian noise variables in the reverse sampling. We also used a pre-trained denoising network as the starting point in the reverse process and fixed the reverse Gaussian noise variables. In addition, we combined both denoising diffusion probabilistic and denoising diffusion implicit models to improve image texture. Lastly, to achieve dose-aware denoising, total injected dose was added as an additional conditional information.

A total of 9,723 ¹⁸F-FDG 3D PET image volumes from 1,586 patients were included. 429 patients were used for network training and validation, and the remaining 1,157 patients (5,883 low-count/low-dose image volumes) were used for network testing. The datasets were collected at Yale New Haven Hospital, Shanghai Ruijin Hospital, and University of Bern Hospital. Three types of different scanners were used in different hospitals. Low-count levels ranging from 1% to 50% were generated through listmode rebinning. 20 real low-dose patient studies were acquired using a United Imaging uExplorer scanner with an average injected dose of 27.1±5.4 MBq. For image quality evaluation, we randomly selected 45 patients from each of the scanner (15 each) and all the 20 real low-dose patients for a reader study (total of 65 patients). Three nuclear medicine physicians from the Yale New Haven Hospital participated in the reader study independently. Multiple 3D images volumes (in DICOM format) of the same patient generated by various methods were provided to the reader each time, and readers were asked to rank each image based on the overall image quality. Readers were also asked to comment on whether any lesions were presented. Images were quantitatively evaluated using SSIM, PSNR, and NRMSE across all the 1,157 testing patients. We also compared DDPET-3D with Unified Noise-aware Network (UNN), a previously method for noise-aware denoising.

Results: DDPET-3D consistently produced superior denoising results to those of UNN for a variety of input count levels. Based on the reader study results, all three readers agree that, at 25% and 50% low-count levels, DDPET-3D produced images with better or similar overall quality to the 100% full-count images across all three scanners. DDPET-3D also consistently received superior ranking scores for the real low-dose scans. DDPET-3D also consistently performed better than UNN quantitatively. For example, at 5% low-count level, the NRMSE values for low-count images, UNN, and DDPET-3D results are 0.245, 0.175, and 0.151, respectively for the United Imaging dataset. These numbers are 0.220, 0.141, and 0.127 for the Siemens Vision Quadra datasets, and 0.474, 0.291, and 0.262 for the Siemens mCT dataset.

Conclusions: Evaluated on a large number of multi-center data with a reader study, DDPET-3D demonstrated its potential to produce high quality low-count PET images.

• Download figure
• Open in new tab
• Download powerpoint

• Download figure
• Open in new tab
• Download powerpoint