Converting Magnetic Resonance Images to Computed Tomographies
Abstract
Purpose
Magnetic resonance imaging (MRI) offers superior soft tissue contrast without ionizing radiation, but computed tomography is necessary for electron density information in radiation therapy treatment planning. This study evaluated the feasibility and generalizability of generating synthetic CT (sCT) images from MRI using a deep learning approach for prostate imaging.
Methods
A two-dimensional U-Net convolutional neural network was trained using paired MRI and CT data from the SynthRAD 2025 prostate dataset. The dataset consisted of 120 T1-weighted (T1W) cases and 60 T2-weighted (T2W) MRI-CT axial pairs, along with an anatomical mask. MRI intensities were normalized between [0,1] and CT values were clipped from -1000 HU to 1500 HU. Synthetic CT slices were generated and stacked to reconstruct three-dimensional 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) across various classes: body contour, fat, muscle, and bone. External validation was performed using Virginia Commonwealth University’s True Fast Imaging with Steady-State Free Precession (TRUFI) MRI and corresponding CT data from ViewRay MRIdian 0.35T MR-Linac System.
Results
The model demonstrated strong agreement between synthetic and ground truth CT images in soft tissue regions. Structural similarity (SSIM) values approached 0.99 for fat and 0.98 for muscle, with mean absolute errors (MAE) of approximately 30 to 40 Hounsfield units. Body contour accuracy was slightly lower, with mean absolute errors of approximately 60 to 70 Hounsfield units. Bone tissue showed higher error, with MAE exceeding 300 Hounsfield units, reflecting known limitations of MRI-based bone modeling. External validation preserved tissue-specific performance trends, supporting model generalizability.
Conclusion
Deep learning-based synthetic CT generation from MRI is feasible for prostate imaging and demonstrates strong performance with soft tissue while cortical bone remains challenging. The results show it is feasible to produce sCT images from MRI data.