Abstract
<title>Abstract</title> <p> <bold>Rationale and Objective</bold> : This study aimed to develop and evaluate a dual-cycle-consistent generative adversarial network (CycleGAN) for synthesizing T1-weighted MR images from brain CT scans, with glioblastoma multiforme (GBM) as the clinical use case. <bold>Materials and Methods</bold> : A prospective dataset of 40 patients was collected, comprising paired and unpaired CT–MRI scans. A CycleGAN framework with adversarial, cycle-consistency, voxel-wise, and identity losses was trained for cross-modality synthesis. Model performance was evaluated using Mean Squared Error (MSE), Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), Fréchet Inception Distance (FID), and binary accuracy. Qualitative analysis was conducted by comparing CT, synthetic MRI, and real MRI images. <bold>Results</bold> : The model achieved an average MSE of 1 0.133 ± 0.017, PSNR of 15.3 ± 0.8 dB, SSIM of 0.60 ± 0.03, FID of 154.4 ± 6.2, and accuracy of 92.4 ± 0.6%. Synthetic MRIs reproduced major neuroanatom-ical landmarks, including ventricular and cortical structures, with improved tissue contrast relative to CT. Clinical reviewers confirmed their plausibility for gross anatomical assessment, though fine structural details and tumor margins remained less precise. <bold>Conclusion</bold> : The proposed CycleGAN framework demonstrates the feasibility of generating synthetic MRI from CT, offering improved soft-tissue visualization and potential utility in GBM (RT) radiotherapy planning. While current performance does not yet replace clinical MRI, the approach provides a foundation for future refinements and holds promise for integration into resource-limited radiotherapy workflows. </p>