Author: Osama R. Mawlawi, Yiran Sun 👨🔬
Affiliation: RICE University, UT MD Anderson Cancer Center 🌍
Purpose: Conventional PET reconstruction methods often produce noisy images with artifacts due to data/model mismatches and inconsistencies. Recently, deep learning-based conditional denoising diffusion probabilistic generative models (cDDPM) [1] have shown great promise in generating realistic results. They however suffer from insufficient correspondence and consistency when conditions and reconstructed images are defined in two different domains (i.e. sinogram versus image space). Building on R2U-DDPM [2], we propose a novel unified framework, Posterior-Mean Diffusion Model (PMDM), which leverages a mean squared error (MSE) supervised pre-trained model (no specific neural network architecture constraint) as an auxiliary prior to enhance cDDPM reliability.
Methods: PMDM first approximates the posterior mean predictions by using a model that minimizes MSE between predictions and ground truth PET images. Then we train a cDDPM to predict the noise at randomly selected timestep between pairs of posterior mean predictions and ground truth PET images. We evaluate the performance of PMDM on simulated paired sinogram and PET images. Specifically, we simulated 2D 18F-FDG PET images using 100 3D brain phantoms from BrainWeb [3]. For each 3D brain phantom, we selected 55 non-continuous axial slices to generate high count sinograms which were then used to reconstruct reference PET images. We used 85 brain samples (4675 slices) for training, 5 brain sample (275 slices) for validation and 10 brain samples (550 slices) for testing. All experiments were implemented in PyTorch and executed on NVIDIA A100 GPUs. The generated PET images were compared to the ground truth as well as those from cDDPM using PSNR (dB)/SSIM.
Results: Generated images are shown in Fig. 1. The average PSNR (dB)/SSIM from PMDM was 27.88/0.956, while from cDDPM was 27.00/0.945.
Conclusion: Our results show that PMDM framework has the potential to produce images that are more qualitatively and quantitatively close to reference images compared to cDDPM.