2241 - Pretreatment Prediction of Multi-Leaf Collimator Positional Deviations in Patient-Specific Plan Control Points Using Complexity Metrics

L. Chen, H. Luo, L. Tan, B. Feng, X. Yang, and F. Jin; Radiation Physics Center, Chongqing University Cancer Hospital, Chongqing, China

Purpose/Objective(s): The accuracy of dose delivery of volumetric-modulated arc therapy (VMAT) plans is influenced by the positional deviations of the multi-leaf collimator (MLC). This study analyzes MLC and gantry positions recorded in trajectory log files, introducing a machine learning model based on plan complexity to predict MLC positional deviations at each control point in patient plans and elucidate the underlying causes of these deviations. Materials/

Methods: Trajectory log files from VMAT plans for 272 cervical cancer patients treated on technology company accelerator and Edge accelerator were collected and analyzed. MLC leaf positions at each control point were calculated based on corresponding monitor unit (MU) interpolation, then, deviations of MLC positions between log file and treatment plan were assessed. A total of 16 control point-level plan complexity metrics, encompassing leaf velocity, acceleration, and sequence variability, were calculated from the patients plan file. The correlation between plan complexity and MLC deviation was examined. Linear and first-order Fourier functions were utilized to fit and establish the prediction model based on linear accelerator motion parameters, with accuracy assessed using the goodness of fit (R²). To enable the prediction with patient plans prior to treatment, probability density distributions of fractionation-dependent model parameters within the same plans were characterized using kernel density estimation.
Results: Deviations between planning and actual MLC positions were up to 3mm on the Clinical IX accelerator and 1.5 mm on the Edge accelerator. A strong linear relationship was observed between MLC leaf positional deviation and leaf velocity, with an average R²=0.87 across all control points in the entire patient plan and R²=0.97 at each single control point. However, the slope and intercept of the fitting varied during the treatment arc. The slope exhibited a high linear correlation with delay time of gantry angle relative to MU (R²=0.92), with probability densities fitted using two mixtures of Gaussian kernels with mean values of 0.14 and 0.06s. While the intercept changed periodically with a 2p period, which may relate to gravity.
Conclusion: This study successfully establishes a control point-level prediction method for MLC positional deviations using machine learning. The method accurately predicts MLC positional deviations during VMAT plan delivery and provides insights into the contributing factors. Accurate prediction of MLC positional deviations holds potential applications in dose calculation and optimization.