S1530
Interdisciplinary - Quality assurance and risk management
ESTRO 2026
Purpose/Objective: Manual RTQA is essential for maintaining protocol compliance in international clinical trials. However, it is resource-intensive and time-consuming, particularly for anatomically complex sites such as gastric cancer and may delay treatment initiation across time zones. Within the TOPGEAR trial1, rigorous pre-treatment RTQA was implemented2. Subsequent secondary expert review identified consistent contouring errors even among cases that had passed RTQA. This highlighted the need for more efficient and consistent RTQA processes that can support large collaborative trials. Building on the prior TOPGEAR QA analysis, this study evaluated how accurately an AI-based QA tool could identify contour deviations compared with human review. Material/Methods: All cases with retrievable clinical target volume (CTV) data were included. As this analysis evaluated a QA method rather than patient outcomes, all contours were included, regardless of whether they represented final treated volumes.An AI-based QA tool, previously developed and validated on a subset of TOPGEAR cases, was applied to the dataset. The QA tool uses auto-segmentation models to estimate uncertainty bands representing the acceptable range for the protocol-defined CTV and evaluates how well clinician-drawn contours fit within these bands to identify potential protocol deviations (classified as pass or violation). Human review, performed using the same TOPGEAR contouring protocol, served as the gold-standard reference.Agreement between automated and human classifications was described using sensitivity, specificity, and overall concordance. Results: A total of 213 CTV contours were analysed. Compared with human review, the automated tool correctly classified 154 of 213 cases (72 %), with a sensitivity of 91% and specificity of 62%. Seven cases (3%) were passed by the tool but flagged as violations on human review, while 52 (24%) were flagged as violations by the tool but passed on human review. Overall accuracy was consistent with the model’s prior test-set performance, supporting its reproducibility when applied to a broader dataset.Tool-flagged violations were predominantly related to under-contouring, consistent with patterns seen on secondary human review.
Conclusion: The AI-based QA tool showed high sensitivity for detecting contour violations, deliberately over-flagging some cases to minimise missed errors. This conservative approach supports its use as a triage aid to focus expert review where most needed. Flagged patterns mirrored those identified on human re- review, suggesting the tool can reflect expert judgement at scale. Integrating systems such as this AI-based contour QA tool into future trials may help improve the consistency and timeliness of RTQA across centres. References: 1. Leong T, Smithers BM, Michael M, Haustermans K, Wong R, Gebski V, O’Connell RL, et al. Preoperative chemoradiotherapy for resectable gastric cancer. N Engl J Med. 2024 Sep 13;391(19):1810-1821. doi:10.1056/NEJMoa2405195. 2. Lukovic J et al. The feasibility of quality assurance in the TOPGEAR international phase 3 clinical trial of neoadjuvant chemoradiation therapy for gastric cancer. Int J Radiat Oncol Biol Phys. 2023;117(4):850-859. doi:10.1016/j.ijrobp.2023.07.012. Keywords: Artificial intelligence, RTQA, contouring
Digital Poster 3533
Immediate response to a ransomware attack in a multi-vendor radiotherapy department in Latin America Marcelo Ribeiro Picioli, Matias Pino Verdugo, Pablo Gonzalez Mella, Hernan Letelier Flores, Filippo Marangoni, Jhonalbert Aponte Arcila, Jose Luis Rodriguez Mongua, Rixy Plata Villamizar, Alvaro Ruiz Plata, Julia Martini Fontana
Made with FlippingBook - Share PDF online