Mesadieu, John Boby2025-08-152025-08-152025-08-15http://hdl.handle.net/10393/50771https://doi.org/10.20381/ruor-31326Positron Emission Tomography (PET) is a nuclear imaging technique that visualizes metabolic activity in the body by detecting gamma rays emitted from a radiotracer. For accurate quantification of radiotracer uptake, attenuation correction (AC) is essential to compensate for the absorption of photons by different tissues. Traditionally, Computed Tomography (CT) provides ideal attenuation maps for PET correction due to its direct relationship with tissue density. In hybrid PET/Magnetic Resonance Imaging (MRI) systems, however, MRI-based attenuation correction is challenging as MRI cannot directly visualize bone structures critical for proper photon attenuation estimation. This thesis addresses this limitation by developing a novel approach for generating synthetic CT images from MRI data for improved PET AC in cardiac imaging, with applications extending beyond attenuation correction to radiation therapy planning and other inter-modal medical image translation tasks. This research makes three key contributions: (1) a CT template and multi-modal registration pipeline to establish spatial correspondence between MR and CT images, (2) a Kernel Ridge Regression (KRR) implementation serving as a baseline, and (3) a Dual Contrast cycleGAN (DC-cycleGAN) model incorporating a novel mu-map loss for bidirectional MR-CT translation. The dataset consisted of paired MR and CT images from 10 subjects, with 8 used for training and 2 for testing. Results demonstrated that the DC-cycleGAN model using in-phase Dixon MR images (GAN-In) achieved superior performance in MR-to-CT translation, with a 39.5% reduction in Mean Absolute Error (MAE) and a 37.1% improvement in Structural Similarity Index compared to the KRR baseline. When applied to PET reconstruction, the synthetic CT-based AC (GAN-AC) achieved substantial improvements: SUVmean bias was reduced by 93% (from 18.85% to 1.27%), SUVmax bias by 89%, Root Mean Square Error by 45%, and MAE by 47% compared to the vendor-provided MR-based AC (MR-AC). The proposed implementation as a web-based application will enable researchers at The Royal's Institute of Mental Health Research to utilize this technology for improved diagnostic precision in PET/MR imaging.enAttribution-NonCommercial-ShareAlike 4.0 Internationalhttp://creativecommons.org/licenses/by-nc-sa/4.0/Positron Emission TomographyMagnetic Resonance ImagingComputed TomographyAttenuation CorrectionSynthetic CTCycle-consitent Generative Adversarial Network (cycleGAN)Inter-Modal Medical Image TranslationPET/MR ImagingPET/CT ImagingThesis