TY - JOUR T1 - PET Image Denoising Using Structural MRI with a Novel Dilated Convolutional Neural Network JF - Journal of Nuclear Medicine JO - J Nucl Med SP - 434 LP - 434 VL - 61 IS - supplement 1 AU - Mario Serrano-Sosa AU - Karl Spuhler AU - Christine DeLorenzo AU - Chuan Huang Y1 - 2020/05/01 UR - http://jnm.snmjournals.org/content/61/supplement_1/434.abstract N2 - 434Objectives: Previous developments of deep learning techniques to denoise PET images have utilized well-known uNet architecture1,2, wherein the uNet like architectures utilize more complex networks, such as Generative Adversarial Network (GAN) and cycle-consistent GAN, as their generators3,4. Previous works have also shown that incorporating structural MRI as an input improves performance 5; this allows for the network to preserve edges in the PET network output. We have recently introduced the dilated CNN (dNet) for PET denoising 6. In this study, we develop a novel PET denoising model by utilizing the dNet architecture with both PET and MRI inputs (dNetPET/MRI) and compare it to three other deep learning models. Methods: All networks were trained using a 2.5D method. dNet combines the skip connections, similar to uNet, but incorporates a dilated convolution 7. Figure 1 shows the dilated convolution kernels used in dNet. We developed dNetPET/MRI to recover full-count images from low-count images and compared it to three other models: dNet with PET only, uNet with PET/MRI and uNet with PET only inputs (dNetPET, uNetPET/MRI, uNetPET, respectively). Given our previous experience with uNet8,9, we initially optimized uNet and set most parameters identical for dNet for a meaningful comparison. Comparison of these models was evaluated through objective imaging metrics: peak signal-to-noise ratio (PSNR), structural similarity (SSIM) index and mean absolute percent error (MAPE). Residual learning was used in all networks. A total of 35 18F-FDG brain PET/MRI studies, with 8,400 slices in total, were acquired and split into training and testing. Each subject was administered between 148-185 MBq (4-5mCi) of 18F-FDG. Listmode data was collected after the injection of 18F-FDG on a Siemens Biograph mMR PET/MRI scanner. Single static PET images were reconstructed from 10-minute emission data (50-60 minutes after injection) using Siemens’ E7tools with ordered subset expectation maximization (OSEM). Low-count PET data (10% count) were generated through Poisson thinning from the full listmode file. MPRAGE images were used alongside reconstructed low-count PET as inputs to the PET/MRI networks. Results: All four networks were successfully trained to synthesize full-count from low-count PET images as shown in Figure 2. Figure 3 shows SSIM, PSNR and MAPE values for the independent testing set evaluated on the four networks compared to low-count. The dNetPET/MRI performed the best across all metrics and performed significantly better than uNetPET/MRI (pSSIM=0.0218, pPSNR=0.0034 , pMAPE= 0.0305, paired t-test). Also, dNetPET performed significantly better than uNetPET (p<0.001 for all metrics, paired t-test). Trend-level improvements were found across all objective metrics in networks using PET/MRI compared to PET only inputs within similar networks (dNetPET/MRI vs. dNetPET and uNetPET/MRI vs. uNetPET), with PET/MRI networks significantly outperforming PET only networks in PSNR and MAPE (Table 1). Conclusions: This is the first work to use dNet architecture for low-count PET denoising with structural MRI as additional input. This network has shown to outperform uNet across various objective imaging metrics. Although various other methods have been introduced recently for low-count PET denoising, such as GAN, cycle-consistent GAN, etc., they typically all use generators that have uNet-like architectures. Combining these novel GAN’s with the expanding field of view that dilated kernels allows may improve upon previously acquired results - further advancing the field of PET/MR denoising. View this table:Table 1: Paired t-test between dNet models and uNet models. ER -