Synthetic CT Validation for MR-Only Head-and-Neck Radiotherapy
Abstract
Purpose
MRI provides advantages for tumor and critical structure delineation in head-and-neck (H&N) radiotherapy. However, H&N MR-only workflows remain challenging due to the large field-of-view and susceptibility artifacts arising from the sinuses, dental hardware, and immobilization devices. The purpose of this work is to generate and assess synthetic CTs (sCTs) to facilitate MR-only H&N radiotherapy workflows.
Methods
Twenty patients with H&N tumors undergoing IMRT (minimum twenty fractions) were included. Treatment sites varied and included the tongue, nasopharynx, thyroid, and oropharynx. Immobilization included thermoplastic masks and bite blocks. Imaging included 120kVp CT (pCTs) and 3D T1-weighted gradient-recalled-echo (GRE) Dixon MRI (TR/TE=5.5/2.46ms, slices=80-90). sCTs (resolution 1.0x1.0mm2, 2mm slice thickness) were generated from MRIs using a 3D convolutional neural network-based prototype (Siemens Healthineers). sCTs were rigidly co-registered to pCTs and resampled to a uniform matrix. Clinical IMRT plans were forward-calculated onto the sCT and compared using gamma analysis and dose-volume-histograms.
Results
For the twenty patients evaluated, average mean absolute error (within the body intersection) was 134.56±15.88 HU. Mean gamma passing rates between plans were 97.83±3.88% (3%/3 mm), 96.14±5.20% (2%/2 mm), and 91.25±7.89% (1%/1 mm). Dose to 95% of PTVs agreed to within 0.15±1.39%. sCTs were accurately reconstructed even for bulky tumors and atypical bone anatomy (e.g. missing maxilla). Failure modes included false high-density material abutting bolus (n=1) and underestimation of bone in the nasal sinus (n=5) and oral hardware (n=2) but yielded negligible dosimetric impact.
Conclusion
This work demonstrates clinical feasibility of sCTs for MR-only H&N radiotherapy. sCTs demonstrated strong agreement with pCTs in terms of image fidelity and dose calculation. Residual errors were primarily associated with challenging patient-specific conditions such as bolus material, hardware, and bone/air cavities. Future work will focus on improving generalization through optimized preprocessing and artifact-aware modeling, including patients with a wider range of immobilization devices.