A fast and reliable quality assurance framework for CBCT-based synthetic CTs for adaptive proton radiotherapy.

Cone beam CT based synthetic CTs (sCTs) enable online adaptive proton therapy by preserving daily patient anatomy without additional acquisition of diagnostic-quality CTs. However, clinical adoption is limited by the lack of fast, reliable quality assurance (QA) methods. This study proposes and evaluates a practical QA workflow to assess sCT suitability for proton dose optimization. Approach: A two-step sCT-QA workflow was developed in which a candidate sCT intended for reoptimization (sCTref) is compared with an independently generated secondary sCT (sCTtest). The workflow combines image based metrics and dose based metrics, with doses derived from recalculating a template proton plan on both sCTs. In this study, the framework was evaluated using a deformably propagated virtual CT (vCT) as sCTref and either a deep learning-based sCT (sCTDL) or an analytically corrected CBCT (CBCTcorr) as sCTtest. Same-day CTs were deformably registered to the CBCT and used as ground truth (CTGT) for evaluation. Thresholds were optimized using 5-fold cross-validation on 182 head and neck (H&N) CT-CBCT pairs. Dose calculation suitability was defined as a gamma pass rate (2%, 1mm) exceeding 95% between dose distributions computed on sCT and CTGT. The method was further tested on an external H&N dataset and on abdominal cases. Main Results: Optimal classification performance was achieved using a combination of image mean absolute error (MAE) and dose MAE. On the H&N test set, the workflow achieved balanced accuracies of 91.2% and 89.7% for the vCT-CBCTcorr and vCT-sCTDL comparisons, respectively, while maintaining a low false positive rate. When applied without modifications to the external H&N and abdominal datasets, balanced accuracy was 94.7% (HN) and 85.0% (abdominal) for vCT-CBCTcorr, and 37.9% (HN) and 61.1% (abdominal) for vCT-sCTDL. Significance: This study introduces a clinically feasible sCT QA workflow for CBCT based adaptive proton therapy and showed promising transferability across datasets.