KEY POINTS
- The study used same-day CT and T1-weighted 3-T MRI from 68 previously treated brain metastasis patients. Fifty cases trained a consistency-regularized generative adversarial network, while 18 independent patients were used to test synthetic CT dosimetry and synthetic digital reconstructed radiographs for CyberKnife image guidance.
- Synthetic CT image quality was high, with mean absolute error 10.08±2.14 HU, structural similarity 0.98±0.01, and peak signal-to-noise ratio 34.73±1.27 dB. Synthetic radiographs achieved SSIM 0.95±0.03, with the largest discrepancies concentrated at skull and nasal bone-air interfaces.
- Auto-contoured large organs were reproduced reasonably well: mean Dice coefficients were 0.88 for brainstem and 0.91 for eyeballs. Smaller structures performed substantially worse, with mean Dice values of 0.60 for lenses and 0.53 for the optic pathway, supporting mandatory manual review.
- When existing CyberKnife plans were recalculated on synthetic CT, target differences versus real-CT plans were very small: V100% changed by −0.05 percentage points, D95% by −0.04 Gy, and mean target dose by −0.06 Gy, with no statistically significant target differences.
- Plans re-optimized directly on synthetic CT also maintained similar target dose: V100% changed by −0.11 percentage points, V95% by −0.12 points, D95% by −0.03 Gy, and mean dose by −0.06 Gy, again without significant target differences.
- The main dosimetric vulnerability appeared in low-dose normal brain. In synthetic-CT-based replanning, normal-brain V3Gy differed by −14.73 cm³ on average, with a very wide confidence interval; the authors linked this partly to the shorter MRI scan range restricting non-coplanar CyberKnife beam geometry rather than synthetic CT accuracy alone.
- Synthetic radiographs produced submillimetre mean translational differences from real-CT radiographs, with no significant between-group differences. Rotational errors were much more dispersed, however, with several outliers exceeding 5°, although mean pitch and yaw remained around 1° and within robotic correction capability.
CLINICAL TAKEAWAY
MR-only CyberKnife treatment for brain metastases appears technically feasible for dose calculation and translational skull tracking. The weak points are small-organ segmentation, bone-air synthesis and rotational registration, so this remains a feasibility workflow requiring prospective QA, full six-degree-of-freedom validation and multicentre testing.