Abstract
437
Objectives: Quantitatively accurate PET images are essential in psychiatric studies to assess subtle differences between diseases. Recently, development of denoising techniques using convolutional neural networks (CNN) for quantitatively accurate low-count PET reconstruction has attracted substantial interest. In this work, we use kinetic modeling to evaluate the performance of our novel multiscale dilated CNN (dNet) and compare it to the well adopted uNet architecture, and, more importantly, full-count data. With dilated kernels, our dNet avoids down-sampling and up-sampling feature maps, a technique used in uNet which typically degrades resolution and fine details. Methods: All models had five channels and were trained using a 2.5D scheme (slabs) to afford the network a degree of 3D information. These slabs contained the slice of interest (middle slice) with two superior and two inferior slices to form a 3D slab. A total of 35 subjects were acquired and split into training (n=30, ~3000 slabs) and testing (n=5). Each subject was administered between 148-185 MBq (4-5mCi) of 18F-FDG and asked to void their bladder immediately after injection. This study acquired listmode data using a dedicated MRI head coil for 60 minutes immediately after the injection of 18F-FDG using a Siemens Biograph mMR PET/MRI scanner. Attenuation maps were generated using an established MRI-based algorithm1,2. Scanner attenuation maps were also extracted for reconstruction. We developed two deep learning models for comparison of PET image denoising: a conventional uNet and our proposed dNet (Figure 1). Kinetic modeling was used to analyze both CNN models. Metabolic rate of glucose (MRglu) was calculated at the voxel-level using the Patlak algorithm3. We used seven frames starting at 10 minutes after injection with temporal frame lengths of 4x5min and 3x10min4. Each frame was reconstructed using OSEM (6 iterations, 21 subsets), interframe motion was corrected and images were then registered to the subject’s T1w MRI. The CNN networks were retrained for each temporal frame using the training sets and methods as listed above. Denoising was performed using the corresponding network for images of a given frame. In the voxel-level analysis, mean absolute percent error (MAPE) of the CNN-denoised data relative to the full-count dataset was the figure of merit. Residual learning was included into both networks.
Results: Quantitative analysis showed both CNNs demonstrate good performance in voxel-level Patlak modeling compared to low-count data. Figure 2 shows a histogram of absolute percent error gathered across all testing set subjects. Both dNet (mean = 8.48%) and uNet (mean=9.61%) significantly outperformed the low-count PET data (mean=12.72%). Representative sagittal and coronal slices of voxel-level MRglu images are shown in Figure 3. Conclusion: Our novel approach of using a dilated convolutional neural network architecture for low-count PET denoising is a quantitatively accurate way to recover full-count PET images. Future denoising machine learning methods could utilize the spatially conserving aspect of dNet to improve their results.
In this issue
Jump to section
Related Articles
Cited By...
- No citing articles found.