PET Image Reconstruction Incorporating Deep Image Prior and a Forward Projection Model

Convolutional neural networks (CNNs) have recently achieved remarkable performance in positron emission tomography (PET) image reconstruction. In particular, CNN-based direct PET reconstruction, which directly generates the reconstructed from sinogram, has potential applicability to enhancements because it does not require reconstruction algorithms, often produce some artifacts. However, these deep learning-based, algorithms disadvantage that they a large number of high-quality training datasets. this study, we propose an unsupervised method incorporates prior framework. Our proposed forward projection model with loss function achieve sinograms. To compare our filtered back (FBP) and maximum likelihood expectation maximization (ML-EM) evaluated using Monte Carlo simulation data brain [$^{18}$F]FDG scans. The results demonstrate quantitatively qualitatively outperforms FBP ML-EM respect peak signal-to-noise ratio structural similarity index.