In-Silico Assessment of Automated CTV Contour Quality Assurance Integration In a Gastric Cancer Clinical Trial
Abstract
Purpose
Radiotherapy quality assurance (RTQA) of target volume delineation is critical in clinical trials to ensure protocol adherence. Manual real-time review is resource-intensive and often limited to a subset of patients. This study evaluates the feasibility of automating contour QA to augment peer review and expand QA coverage across the entire trial cohort.
Methods
An in-silico trial was simulated using 203 cases from the AGITG/TROG TOPGEAR gastric cancer trial. Auto-segmentation models were trained using a Probabilistic U-Net to estimate uncertainty bands reflecting the acceptable range of the Clinical Target Volume (CTV). Metrics describing the fit of the manual contour to the uncertainty band (distance-to-band, volume difference, band overlap fraction) were extracted and used to train logistic regression classifiers. These classifiers were trained on successive batches of 10 cases to simulate a prospective trial rollout, iteratively updating the model while tuning for high sensitivity to minimize false negatives.
Results
Across 193 prospective simulation cases, the automated QA model achieved a sensitivity of 0.89, correctly flagging 65 of 73 violations for review. A distinct learning curve was observed, with lower sensitivity (0.63) in the initial 20 cases before the model stabilized to consistent performance as it adapted to the trial distribution. A specificity of 0.54 was observed, with 65 of 120 acceptable contours correctly passed. Implementation of this triage mechanism would have resulted in a 38% reduction in manual QA workload (73/193 cases passed), reducing the need for manual review while maintaining high sensitivity for detecting protocol violations.
Conclusion
Automated contour QA can feasibly streamline RTQA in clinical trials. Applied to the TOPGEAR trial, our approach demonstrated the potential to significantly reduce manual review workload. Crucially, the in-silico simulation exhibited an initial convergence phase, indicating the need for supervised operation before safe deployment of autonomous triage in multi-center trials.