Full‐count PET recovery from low‐count image using a dilated convolutional neural network

Abstract

Purpose: Positron emission tomography (PET) is an essential technique in many clinical applications that allows for quantitative imaging at the molecular level. This study aims to develop a denoising method using a novel dilated convolutional neural network (CNN) to recover full-count images from low-count images. Methods: We adopted similar hierarchical structures as the conventional U-Net and incorporated dilated kernels in each convolution to allow the network to observe larger, more robust features within the image without the requirement of downsampling and upsampling internal representations. Our dNet was trained alongside a U-Net for comparison. Both models were evaluated using a leaveone-out cross-validation procedure on a dataset of 35 subjects (~3500 slabs), which were obtained from an ongoing 18 F-Fluorodeoxyglucose (FDG) study. Low-count PET data (10% count) were generated by randomly selecting one-tenth of all events in the associated listmode file. Analysis was done on the static image from the last 10 minutes of emission data. Both low-count PET and full-count PET were reconstructed using ordered subset expectation maximization (OSEM). Objective image quality metrics, including mean absolute percent error (MAPE), peak signal-to-noise ratio (PSNR), and structural similarity index metric (SSIM), were used to analyze the deep learning methods. Both deep learning methods were further compared to a traditional Gaussian filtering method. Further, region of interest (ROI) quantitative analysis was also used to compare U-Net and dNet architectures. Results: Both the U-Net and our proposed network were successfully trained to synthesize fullcount PET images from the generated low-count PET images. Compared to low-count PET and Gaussian filtering, both deep learning methods improved MAPE, PSNR, and SSIM. Our dNet also systematically outperformed U-Net on all three metrics (MAPE: 4.99 AE 0.68 vs 5.31 AE 0.76, P < 0.01; PSNR: 31.55 AE 1.31 dB vs 31.05 AE 1.39, P < 0.01; SSIM: 0.9513 AE 0.0154 vs 0.9447 AE 0.0178, P < 0.01). ROI quantification showed greater quantitative improvements using dNet over U-Net. Conclusion: This study proposed a novel approach of using dilated convolutions for recovering fullcount PET images from low-count PET images.

Multitask Learning Based Three‐Dimensional Striatal Segmentation of MRI: fMRI and PET Objective Assessments

“…38 Likewise, a recent network architecture has been proposed that replaces down/upsampling in encoder-decoder architectures with dilated convolutions (dNet). 39 The consistent change in image resolution from down/upsampling layers leads to a degree of blurring, which may result in smoothed edges. Therefore, incorporating the dNet architecture, which utilizes dilated kernels to increase receptive field size, may be of use to alleviate the problem of smoothed edges.…”

“…Disparity estimation is a method that uses weighted average operation on the probability distribution to determine segmentations; this method has been shown to alleviate the problems of preserving local structure of edge boundaries 38 . Likewise, a recent network architecture has been proposed that replaces down/upsampling in encoder‐decoder architectures with dilated convolutions (dNet) 39 . The consistent change in image resolution from down/upsampling layers leads to a degree of blurring, which may result in smoothed edges.…”

Deep learning-based image reconstruction and post-processing methods in positron emission tomography for low-dose imaging and resolution enhancement

“…Sanaat et al [ 80 ] compared low-dose to full-dose mappings with a 3D Unet in image space and sinogram space and demonstrated that sinogram space processing produced improved results with significantly higher PSNR and significantly lower SUV bias. Spuhler et al [ 79 ] used a CNN formulation with dilation of the convolutional kernels in place of down-sampling operations, with PSNR, SSIM and NRMSE results comparable to a Unet. Wang et al [ 81 ] trained a feed-forward fully connected network with synthetic data to denoise low-dose 4 × 4 × 4 pixel patches and directly applied the network to 82 Rb cardiac PET images.…”

Applications of artificial intelligence in nuclear medicine image generation

“…Such a network lacks sufficient knowledge to distinguish noise from useful information. Adjacent slices can be used as different input channels [PET-plus-MR or PET-only image input (165,166), three-layer PET image input (167,168), and five-layer SPECT image input (124,125)] can provide 2.5D structural information to the network, and this can be called a method of target feature enhancement. Compared with 3D convolution, calculation costs are high for this approach.…”