Research on real-time detection of radiotherapy setup errors and intelligent quality control methods based on...

Area of Science:

• Medical Physics
• Machine Learning in Healthcare
• Radiotherapy Quality Assurance

Background:

• Accurate patient positioning is critical for effective radiotherapy.
• Current methods for detecting setup errors may not provide timely alerts.
• Developing automated systems for real-time error detection is essential for quality assurance.

Purpose of the Study:

• To develop and validate an unsupervised machine learning (ML) model for near-real-time detection of statistically abnormal six-dimensional (6D) radiotherapy setup errors.
• To evaluate the robustness of the ML approach across various immobilization techniques and treatment sites.
• To establish a "setup-monitoring-alert" framework to enhance radiotherapy quality assurance.

Main Methods:

• Analysis of 7,539 CBCT-based 6D setup error records.
• Development and comparison of two unsupervised ML models: Isolation Forest (IF) and Local Outlier Factor (LOF).
• Performance evaluation using ROC-AUC, PR-AUC, and sensitivity at a fixed false positive rate (FPR ≈ 5%), with subgroup analyses by immobilization method and treatment site.

Main Results:

• The Isolation Forest (IF) model demonstrated superior performance (ROC-AUC = 0.960) compared to LOF (ROC-AUC = 0.880).
• High performance (AUC ≥ 0.92) was achieved across most immobilization methods, with specific mask combinations showing near-ideal results.
• Interpretability analysis identified AP, Pitch, and LR directions as key contributors to abnormality detection, with stable performance over time.

Conclusions:

• Unsupervised learning is feasible for identifying statistically unusual radiotherapy setup patterns.
• A closed-loop "setup-monitoring-alert" framework can be implemented to support clinical workflows.
• This ML-based approach serves as an auxiliary alerting tool, complementing dosimetric evaluation and clinical decision-making.