MRI-only Radiotherapy Dose Planning via CycleGAN-Generated Synthetic CT

Abstract

Magnetic resonance imaging (MRI)-only radiotherapy planning seeks to replace computed tomography (CT) by generating synthetic CT (sCT) images directly from MRI, exploiting MRI’s superior soft-tissue contrast; however, MRI lacks the electron density information required for accurate dose calculation, necessitating a dual-modality CT–MRI workflow that increases scanning time and registration uncertainty. This CT–MRI paradigm subjects patients to additional radiation and prolonged imaging sessions, which can degrade planning accuracy, making a reliable MRI-only solution critical for safer, faster, and more precise radiotherapy. To address this, an end-to-end CycleGAN framework is presented to synthesize CT images from routine T1-weighted brain MRI using unpaired data, eliminating the need for exact MRI–CT pairs; the architecture employs U-Net-based generators and PatchGAN discriminators with cycle-consistency and identity losses for robust domain translation. On 100 held-out paired MR–CT slices, the generated sCT achieved a mean absolute error of 58 ± 10 HU and a structural similarity index of 0.92 ± 0.03 compared to ground-truth CT, preserving bone interfaces, air cavities, and soft-tissue boundaries, thus demonstrating suitability for dosimetric integration.