Risk stratification and cross-center prediction of thoracoabdominal 6D setup errors.

Area of Science:

• Radiation Oncology
• Medical Imaging
• Machine Learning in Healthcare

Background:

• Six-degree-of-freedom (6DoF) setup errors in Cone-Beam Computed Tomography (CBCT)-guided radiotherapy vary significantly between treatment centers.
• Accurate patient positioning is crucial for effective radiotherapy delivery and minimizing off-target radiation exposure.

Purpose of the Study:

• To investigate if pre-treatment workflow variables can stratify the risk of large translational setup errors (≥5 mm) derived from CBCT.
• To evaluate the performance of these variables for predicting infrequent rotational errors (≥3°) and assess model generalizability across different centers.

Main Methods:

• A retrospective analysis of 6,033 CBCT-derived 6DoF setup records from two centers was performed.
• Machine learning models (CatBoost, XGBoost) were trained and validated using pre-treatment variables, assessing endpoints like translational displacement (Y5mm, Y10mm) and rotational magnitude (R3°).
• Bidirectional cross-center validation was employed to test model transferability and robustness.

Main Results:

• The study identified significant differences in setup error rates between centers, indicating domain shift.
• Machine learning models demonstrated modest predictive performance for large translational errors (Y5mm), with limited success for rotational errors (R3°).
• Cross-center validation revealed performance degradation and asymmetric transfer patterns, highlighting the need for local model adaptation.

Conclusions:

• Pre-treatment workflow variables can provide some risk stratification for translational setup errors in CBCT-guided radiotherapy.
• The predictive utility for rotational errors is limited, and models exhibit substantial heterogeneity across centers.
• Local recalibration of predictive models at the treatment center is recommended for optimal workflow integration and reliable error prediction.