Development of an MR-only radiotherapy treatment planning workflow using a commercial synthetic CT generator for brain and head & neck tumor patients.

Background: In magnetic resonance (MR)-only radiotherapy (RT) workflows, synthetic computed tomography images (sCT) are needed as a surrogate for a dose calculation. Commercial and certified sCT algorithms became recently available, but many have not been evaluated in a clinical setting, especially in the head and neck tumor (HN) region. In this study, an MRI-only workflow using a commercial sCT generator for photon beam therapy in brain and HN body sites was evaluated in terms of dose calculation accuracy, modelling of immobilization devices, as well as usability for autosegmentation.

Methods: For 13 brain and 10 HN cancer patients, MR scans using T1W mDIXON sequences were retrospectively collected. Four brain and all HN patients were scanned in RT treatment position with immobilization devices. All MRIs were converted to a sCT using the MRCAT algorithm (Philips, Eindhoven, The Netherlands). All patients underwent standard planning CT (pCT) for clinical segmentation and VMAT treatment planning. The sCT was rigidly registered to the pCT and clinical contours were transferred to the sCT. For dosimetric evaluation of sCT based dose calculation, all VMAT plans were recalculated on the sCT. D and D were compared for all structures between pCT and sCT, but D, D for targets only. For MR-invisible RT immobilization device modelling, MR-visible markers were placed into sCT and a geometric robustness analysis was performed based on the same target dose-volume parameters. For organs-at-risk (OARs) autosegmentation, both pCT and sCT were autosegmented with a clinically established CT-based autocontouring software. The agreement of contours on pCT and sCT was analyzed by similar dose-volume parameters and dice similarity (DSC) and Hausforff distance (HD).

Results: The overall median deviation (± interquartile range) of dosimetric parameters between sCT and pCT including the immobilization model was 1.1 ± 0.4% for brain target volumes, 1.3 ± 1.2% for brain OAR, 0.4 ± 0.7% for HN target volumes and 0.4 ± 0.9% for HN OAR. The median geometric agreement over all sCT autocontours compared to pCT autocontours resulted in DSC = 0.82 for brain OAR and DSC = 0.79 for HN OAR.

Conclusion: MR-only RT planning using MRCAT software package was feasible for brain and HN tumors, with acceptable clinical accuracy. The MR-invisible immobilization devices could be modelled in the planning system and the autosegmentation on sCTs using a CT-based autosegmentation tool was feasible.