engleski

PET image reconstruction - a theoretical overview with deep learning and compressive sensing

PET imaging is a medical modality typical used in multimodal imaging in PET/CT or PET/MR. Two types of standard reconstruction methods are discussed - analytical (filtered backprojection) and iterative (maximum likelihood expectation maximization). Analytical reconstruction does not incorporate the system matrix (which models the PET scanner) in reconstruction and, therefore, produces images of lesser quality. Contrary, iterative reconstruction models the PET scanner by the system matrix. Reconstructed images with iterative methods are better when compared to analytical, but the system matrix can be hard to calculate. In recent years, neural networks were introduced to PET imaging. Neural networks can speed up the calculation of the PET scanner model and can enhance standard methods. Additionally, deep learning can be used to suppress the noise in reconstructed images and can be exploited in image reconstruction. The compressive sensing approach is employed to reduce the number of measurements and, therefore, make acquisition time shorter. The compressive sensing approach needs to be investigate further.

positron emission tomography ; image reconstruction ; neural networks ; deep learning ; compressive sensing

nije evidentirano