Deep learning-based denoising for PennPET Explorer data

Objectives: Promising results have been reported for low-statistics PET data denoising using a deep convolutional neural network (CNN) with standard-statistics images as the target and low-statistics images as input for training. The trained network can generate denoised low-statistics images, which have similar noise level as the standard-statistics images. The long axial field-of-view (AFOV) scanner of the PennPET Explorer leads to very high sensitivity and permits low dose, fast, and/or late imaging. A CNN trained on the high-count data from the PennPET scanner will be able to generate virtual-high-statistics images to permit ultra-low dose, ultra-fast, and/or ultra-late imaging. In this study, we investigated and optimized a fully 3D U-Net architecture for the PennPET Explorer data denoising. Methods: Two subjects were injected with 15 mCi FDG and scanned in a single bed position on the time-of-flight (TOF) PennPET Explorer scanner (currently with 64-cm AFOV). Subject #1 was scanned for 20 min after 105 min post-injection. Subject #2 was scanned for 20 min after 85 min post-injection, and then an ultra-late scan was acquired for 60 min after 10 half-lives of FDG (~18 hr post-injection). For each subject, 10 low-count (5% of full count, 1-min data) samples were generated by independent sampling from the full-count (20-min) data. All the images were reconstructed by list-mode TOF OSEM (25 subsets, 4 iterations). The image size was 288×288×320 with voxel size of 2×2×2 mm3, and then cropped to 144×272×320 for denoising. A fully 3D U-Net was trained for denoising by minimizing the L2 loss function using Adam optimization. The 10 low-count and 1 full-count images from subject #1 were used as input and target for training, respectively, and the 10 low-count images from subject #2 were used for testing. To augment the training dataset, a patch-based method was used with 64×64×16 patch size. To reduce artifacts on the tile edge caused by the overlapping-tile strategy, 144×272×160 patch size was used in testing. Regions of interest (ROIs) of cerebellum, bone marrow, myocardium, aorta wall, and lung were drawn on the full-count image of subject #2. The quantification accuracy was evaluated using the relative bias of ROI mean value with the full-count image as ground truth. The background noise was evaluated using the normalized standard deviation in the lung ROI. Finally, the trained network was applied to the ultra-late scan of subject #2 to generate a low-noise