PET Image Reconstruction Incorporating Deep Image Prior and a Forward Projection Model
Fumio Hashimoto, Kibo Ote, Yuya Onishi
Abstract
Convolutional neural networks (CNNs) have recently achieved remarkable performance in positron emission tomography (PET) image reconstruction. In particular, CNN-based PET image reconstruction, which directly generates the reconstructed image from a sinogram, has potential applicability in PET image enhancement because it does not require image reconstruction algorithms, which often produce artifacts. However, these deep learning-based PET image reconstruction algorithms have the disadvantage that they require a large number of high-quality training datasets. In this study, we propose an unsupervised PET image reconstruction method that incorporates a deep image prior (DIP) framework. Our proposed method incorporates a forward projection model with a loss function to achieve unsupervised PET image reconstruction from sinograms. To compare our proposed image reconstruction method with filtered back projection (FBP), maximum-likelihood expectation–maximization (ML-EM), and the other DIP-based reconstruction algorithm, we evaluated our method using Monte Carlo simulation data of a brain [18F]fluoro-2-deoxy-D-glucose (FDG) PET scan and real data of a rhesus monkey brain [18F]FDG PET scan. The results demonstrate that our proposed image reconstruction method quantitatively and qualitatively outperforms the FBP and ML-EM algorithms; furthermore, it showed comparable performance and faster calculation time compared to the other DIP-based image reconstruction method.