Related Experiment Video
Updated: Feb 24, 2026
Author Spotlight: Improving Radiation Therapy Access with Radiation Planning Assistant
Published on: October 6, 2023
Research on real-time detection of radiotherapy setup errors and intelligent quality control methods based on
Weixiang Lin1, Chengjian Xiao1, Liangjie Xiao2
1Department of Radiation Oncology, Ganzhou Cancer Hospital, Ganzhou, China.
Objective:
This study aimed to develop and validate an unsupervised machine learning-based approach for near-real-time alerting of statistically abnormal six-dimensional (6D) radiotherapy setup errors. Using large-scale clinical datasets, the robustness of the proposed approach was evaluated across different immobilization methods and treatment sites to support quality assurance alerting.
Methods:
A total of 7,539 CBCT-based 6D setup error records collected at our center between May 2022 and March 2025 were analyzed. After data standardization and construction of proxy anomalous samples, two unsupervised models-Isolation Forest (IF) and Local Outlier Factor (LOF)-were developed. Model performance was assessed using ROC-AUC, PR-AUC, and sensitivity at a fixed false positive rate (FPR ≈ 5%). Subgroup analyses were performed by immobilization method and treatment site. Interpretability was explored using principal component analysis (PCA) and Spearman correlation. To provide minimal translational context, geometric tolerance exceedance rates based on translational and rotational thresholds were quantified.
Results:
Overall, IF outperformed LOF (ROC-AUC = 0.960 [95% CI: 0.956-0.964] vs. 0.880 [95% CI: 0.872-0.888]). Most immobilization methods achieved AUC ≥ 0.92 (range: 0.912-1.000), with dual-face SRT masks and neck-thorax mask plus vacuum cushion combinations approaching ideal performance (AUC ≈ 1.00). Interpretability analyses indicated that the AP, Pitch, and LR directions were the primary contributors to abnormality detection. Longitudinal evaluation revealed stable performance without model drift.
Conclusion:
This study demonstrates the feasibility of applying unsupervised learning to identify statistically unusual setup patterns and proposes a closed-loop "setup-monitoring-alert" framework. The approach is intended as an auxiliary alerting tool to support clinical workflows, rather than to replace dosimetric evaluation or clinical decision-making.
