Dynamic PET Image Denoising with Deep Learning-Based Joint Filtering

Dynamic positron emission tomography (PET) imaging usually suffers from high statistical noise due to low counts of the short frames. This study aims to improve the image quality of the short frames by utilizing information from other modality. We develop a deep learning-based joint filtering framework for simultaneously incorporating information from longer acquisition PET frames and high-resolution magnetic resonance (MR) images into the short frames. The network inputs are noisy PET images and corresponding MR images while the outputs are linear coefficients of spatially variant linear representation model. The composite of all dynamic frames is used as training label in each sample, and it is down-sampled to 1/10th of counts as the training input. L1-norm combined with two gradient-based regularizations constitute the loss function during training. Ten realistic dynamic PET/MR phantoms based on BrainWeb are used for pre-training and eleven clinical subjects from Alzheimer’s Disease Neuroimaging Initiative further for fine-tuning. Simulation results show that the proposed method can reduce the statistical noise while preserving image details and achieve quantitative enhancements compared with Gaussian, guided filter, and convolutional neural network trained with the mean squared error. The clinical results perform better than others in terms of the mean activity and standard deviation. All of the results indicate that the proposed deep learning-based joint filtering framework is of great potential for dynamic PET image denoising.