Evaluation of CBCT-based HU correction for synthetic CT generation in head-and-neck radiotherapy: FOV-stratified and peripheral-region analysis

Background Cone-beam computed tomography (CBCT) is essential for image-guided radiotherapy for head and neck (H&N) cancer but suffers from inaccurate Hounsfield units (HUs) due to scatter, artifacts, and field of view (FOV) truncation, limiting its use for dose calculation. We aimed to compare linear correction, a residual U-Net model (ResUNet), and a paired cycle-consistent conditional GAN (PCC-cGAN) framework with residual U-Net generators for HU correction and synthetic CT (sCT) generation, the deep learning models trained in a paired setting, with the CycleGAN variant adapted from its original unpaired design to exploit cycle-consistency in addition to supervised training. Our evaluation emphasizes FOV stratification (complete vs. truncated) and spatial localization of residual errors. This study provides an FOV-stratified and peripheral-aware evaluation of CBCT-based HU correction, highlighting residual error patterns that may be overlooked by global MAE analysis alone. Methods A total of 100 CBCT/CT pairs from 67 H&N patients were divided into complete FOV (19 pairs) and truncated FOV (81 pairs) groups. Evaluation was performed within three spatial mask regions: intersection, eroded intersection, and peripheral. An independent-patient sensitivity analysis was conducted to assess patient‑level overlap. Preliminary body‑level dosimetric consistency was evaluated by comparing raw CBCT and ResUNet sCT against the reference planning CT (pCT) dose using gamma pass rates (2%/2mm and 3%/3mm) and dose Mean Absolute Error & Root Mean Squared Error (MAE/RMSE). Results ResUNet achieved the lowest HU-MAE across the evaluated regions in both FOV groups. In the independent-patient subset, this pattern was preserved for intersection and eroded mask regions, with ResUNet performing best in nearly all of the patient-region combinations. The peripheral MAE remained substantially higher than the intersection and eroded MAE for all methods (e.g., ResUNet: 171.51 vs. 76.93 HU in the complete FOV). Body‑level gamma pass rates (GPR) and dose metrics were generally maintained or modestly improved by ResUNet, but HU improvements did not translate proportionally to dosimetric gains across all patients. Conclusions ResUNet provided the lowest HU-MAE among the evaluated methods under the conditions of this study for CBCT-based sCT generation in H&N radiotherapy, with persistent residual errors concentrated at the peripheral evaluation region, where the highest residual MAE was observed. The FOV-stratified and peripheral-aware evaluation framework introduced here informs future DVH-based validation and dose-of-the-day adaptive workflows.