부산대학교 도서관

Positron emission tomography (PET) allows an unparalleled insight into a wide range of medical conditions including neurological disorders, oncology and cardiovascular diseases. Despite PET’s clinical utility, further improvements can still be made by tackling the susceptibility of reconstructed PET images to noise, particularly for low count PET projection data. In addition, the limited spatial resolution of the PET system leads to partial volume effects (PVEs), which introduce bias in regional quantification. The recent advent of clinically available simultaneous PET-MR scanners enables the simultaneous acquisition of complementary information within a clinically practical workflow. Reparameterising the PET reconstruction process through algorithms such as the machine learning inspired kernel method (KEM), utilises the readily available MR information and has been shown to aid in the suppression of noise and PVEs. However, uncertainty still persists in the capability of MR-informed methods to handle inconsistencies between the PET-MR datasets. Therefore, this thesis aims to demonstrate the major denoising power provided by the MR-informed kernel method, to show the effectiveness of modified KEM implementations in addressing its limitations, and to assess the performance of multiple KEM methods relative to other state-of-the-art MR-informed methods. An additional research direction, also originating from the machine learning field, is the inclusion of convolutional neural networks (CNNs) in PET reconstruction as a powerful means for image denoising. Further investigation and development is presented for unrolled neural networks, and their performance relative to traditional PET reconstruction methodologies. This thesis therefore identifies the key benefits presented by the inclusion of MR-information into the PET reconstruction process, highlights how best to include MR into regularised or reparametrised PET image reconstruction, and explores the potential advantages of incorporating CNNs into PET reconstruction.