TU-G-BRD-07: A Statistical Learning Approach to the Accurate Prediction of MLC Positional Errors During VMAT Delivery

Authors

  • Carlson J,

    1. Seoul National University, Seoul
    2. Seoul National University Hospital, Seoul
    3. Jeju National University Hospital, Jeju
    4. Seoul Veterans Hospital, Seoul
    Search for more papers by this author
  • Park J,

    1. Seoul National University, Seoul
    2. Seoul National University Hospital, Seoul
    3. Jeju National University Hospital, Jeju
    4. Seoul Veterans Hospital, Seoul
    Search for more papers by this author
  • Park S,

    1. Seoul National University, Seoul
    2. Seoul National University Hospital, Seoul
    3. Jeju National University Hospital, Jeju
    4. Seoul Veterans Hospital, Seoul
    Search for more papers by this author
  • Park J,

    1. Seoul National University, Seoul
    2. Seoul National University Hospital, Seoul
    3. Jeju National University Hospital, Jeju
    4. Seoul Veterans Hospital, Seoul
    Search for more papers by this author
  • Choi Y,

    1. Seoul National University, Seoul
    2. Seoul National University Hospital, Seoul
    3. Jeju National University Hospital, Jeju
    4. Seoul Veterans Hospital, Seoul
    Search for more papers by this author
  • Ye S

    1. Seoul National University, Seoul
    2. Seoul National University Hospital, Seoul
    3. Jeju National University Hospital, Jeju
    4. Seoul Veterans Hospital, Seoul
    Search for more papers by this author

Abstract

Purpose:

To quantify and predict the magnitude of multi-leaf collimator (MLC) positional errors in volumetric modulated arc therapy (VMAT) plans using statistical learning techniques to allow more accurate representation of the dose distribution expected to be delivered.

Methods:

A total of 74 VMAT plans used for patient treatments from three separate institutions were acquired. All plans were delivered using a Varian Millennium 120 MLC. The plans were split into training (N=3), validation (N=6) and testing (N=65) sets. From these, numerical features such as individual leaf position and velocity, and categorical features such as whether the leaf was moving towards or away from the isocenter, the bank the leaf was a part of, and the control point (CP) at which the error occurred were extracted. The differences between planned and delivered leaf positions in the training data were used as a target response for the development of a linear regression model, a decision tree model, and a random forest model. Optimized model parameters were found using cross-validation on the validation set. Performance of each model in predicting the positional errors was assessed using mean absolute error (MAE) and root mean square error (RMSE) on the held-out test set.

Results:

The MAE between planned and delivered positions for moving MLCs was 1.27 mm (RMSE = 1.60 mm). The decision tree model had the best performance, the predictions of which had MAE for moving MLCs of 0.27 mm (RMSE = 0.39 mm). Leaf velocity significantly predicted position errors, (β = 0.128, p<0.0001), and explained a significant amount of the variance (r2=0.90, p<0.0001).

Conclusion:

The decision tree model accurately predicted actual MLC leaf positions during delivery. Incorporating predicted errors into the planned MLC positions leads to a more realistic representation of the leaf locations which can be expected during treatment delivery to the patient.

Ancillary