Defense Date

2026

Document Type

Thesis

Degree Name

Master of Science

Department

Radiation Oncology

First Advisor

Dr. William Song

Abstract

Magnetic resonance imaging (MRI) provides excellent soft tissue contrast without ionizing radiation, making it a strong alternative to computed tomography (CT) in medical imaging workflows. However, CT remains essential for applications requiring electron density information, such as radiation therapy treatment planning. This study investigated the feasibility of generating synthetic CT (sCT) images from MRI using a deep learning-based U-Net architecture. A two-dimensional U-Net was trained on paired MRI-CT data from the SynthRAD 2025 dataset, consisting of 120 T1-weighted (T1W) and 60 T2-weighted (T2W) axial cases, including deformed CT (dCT) aligned to MRI. Model testing used an independent dataset of 30 patients from Virginia Commonwealth University (VCU), consisting of axial TRUFI sequences acquired on the MRIdian system. MRI intensities were normalized to [0,1], and CT values were clipped to −1000 to 1500 Hounsfield units (HU). Slice-wise predictions were stacked to reconstruct three-dimensional sCT volumes. Model performance was evaluated using structural similarity index measure (SSIM), mean absolute error (MAE), peak signal-to-noise ratio (PSNR), and normalized cross-correlation (NCC). Strong performance was observed in soft tissue regions, with SSIM ~0.99 (fat) and ~0.98 (muscle), MAE ~30–40 HU, and PSNR ~30–34 dB. Body contour performance was slightly lower (SSIM ~0.88–0.90, MAE ~60–70 HU). Bone remained the most challenging, with higher error (MAE ~310–330 HU, PSNR ~15–16 dB). External validation preserved these trends, supporting model generalizability. Overall, deep learning-based sCT generation from MRI is feasible and promising, with accurate soft tissue prediction and ongoing need for improved bone modeling.

Rights

© The Author

Is Part Of

VCU University Archives

Is Part Of

VCU Theses and Dissertations

Date of Submission

4-24-2026

Download

Share

COinS