KEY POINTS
- 28 patients underwent a dedicated brain MR-simulation protocol in addition to standard planning CT, using four MRI sequences: T1 VIBE Dixon, T2 SPACE, PETRA and TOF-MRA. Synthetic CT was successfully generated in 25/28 patients; one failure resulted from coil connection failure, one from data loss, and one patient withdrew.
- Inter-sequence motion affected 9/25 datasets (36%) and required rigid re-registration when rotational misalignment exceeded 1°. The authors attributed this partly to the long multi-sequence workflow and the temporal separation between synthetic CT acquisition and post-contrast imaging used for delineation.
- External localization markers interfered with synthetic CT generation in 10/25 cases (40%); signal perturbation occurred in 10/10 patients in whom lipid-filled MR markers were used. Suboptimal receiver-coil placement caused major signal non-uniformity and erroneous classification of parts of the skin and skull as air in 1/25 case.
- Algorithm-related HU misclassification at bone/tissue/air interfaces occurred in 16/25 cases (64%), making it the most common problem. Typical failures included air cavities near the sinuses being reconstructed as bone and underestimation of skull HU; example cases showed corresponding displacement of isodose distributions and reduced target coverage.
- Postoperative anatomy was another vulnerability: 2/5 postoperative patients showed substantial synthetic CT perturbation around surgical meshes, with adjacent bone sometimes classified as air and mesh classified as bone. The authors considered altered anatomy a patient-selection issue rather than something that could necessarily be fixed by workflow changes.
- The study separates problems that can potentially be mitigated operationally, such as patient motion, coil placement and marker use, from intrinsic reconstruction problems requiring case-specific review. It provides dedicated commissioning and per-patient QA checklists covering eligibility, acquisition, physics review, system changes and end-to-end testing.
- A major limitation is that the evaluated reconstruction was atlas-based and has since been superseded by a vendor deep-learning system using a single T1 VIBE Dixon sequence. The authors nevertheless argue that several workflow and abnormal-anatomy failure modes remain relevant regardless of reconstruction architecture.
CLINICAL TAKEAWAY
Controlled dosimetric equivalence studies can make MR-only radiotherapy look simpler than it is. In actual implementation, motion, coil geometry, markers, abnormal anatomy and bone-air classification created frequent failure modes, supporting mandatory patient-specific physics review rather than blind reliance on generated synthetic CT.