Abstract

Magnetic Resonance Imaging (MRI) offers excellent soft-tissue contrast without ionizing radiation but lacks the quantitative attenuation information needed for applications such as radiotherapy planning and PET/MRI attenuation correction. This thesis introduces MAGNet, a mask-guided MR-to-CT synthesis framework that incorporates anatomy-aware priors into a Generative Adversarial Network (GAN) to generate high-fidelity synthetic CT (sCT) from pelvic MRI.

MAGNet uses segmentation masks automatically produced by TotalSegmentator to divide the translation into anatomically informed sub-tasks. Two specialized branches, conditioned on bone and soft-tissue masks, allow the network to focus on thin cortical structures, trabecular detail, and subtle density variations while preserving soft-tissue realism and suppressing artifacts. Outputs from these branches are adaptively fused through a learned blending mechanism, enhancing fidelity at bone-soft-tissue interfaces and in regions prone to motion or susceptibility distortions.

The framework is evaluated on a public pelvic MRI-CT dataset and a multi-scanner internal cohort, and benchmarked against representative learning-based baselines. MAGNet achieves consistently higher PSNR and SSIM, lower mean absolute error, and superior bone and soft-tissue reconstruction quality.

Requiring no manual contouring and adding minimal preprocessing, MAGNet is suited for integration into clinical workflows. By improving the structural fidelity and robustness of MR-to-CT translation, it advances MRI-only imaging pipelines toward reliable, radiation-free substitutes compatible with downstream planning, image-guided interventions, and quantitative analysis.