Predicting Ion Chamber Disagreements for Patient-Specific Quality Assurance of VMAT Plans Using Machine Learning
P Wall*, M Lafreniere, E Hirata, O Morin, G Valdes, A Witztum, University of California San Francisco, San Francisco, CA
Presentations
WE-C-TRACK 3-6 (Wednesday, 7/28/2021) 1:00 PM - 2:00 PM [Eastern Time (GMT-4)]
Purpose: To develop and validate a machine learning model for predicting ion chamber disagreements used in patient-specific quality assurance (QA) of VMAT plans at a multi-site institution.
Methods: A dataset of 579 VMAT treatment plans and associated QA measurements was retrospectively collected at our institution. 30 classes of complexity features representing different failure modes (totaling 293 raw features) were extracted from each DICOM RT-Plan file using in-house software. QA outcomes, represented by the percent difference between planned dose and ion chamber measured dose, were extracted from a RedCAP database used in our clinic. An XGBoost model was trained (10-fold cross-validation) to predict ion chamber measurement disagreements using the complexity features as input. A grid search over a range of values was used to optimize maximum tree depth and learning rate hyperparameters, after which the tuned model was tested on a previously unseen dataset.
Results: The beam model, edge metric, monitor unit factor, small aperture score at 25 mm, and modulation complexity score were found to be the five most important features for predicting ion chamber disagreements by total gain. The initial model was able to predict all but 24 of the 579 testing samples within ±3% across the 10 testing folds. The average training and testing mean absolute error (MAE) were 0.73 ± 0.14% and 1.10 ± 0.17%. The 24 samples predicted outside of 3% error were remeasured after which the model was retrained and reoptimized. All testing samples were subsequently predicted within 3% error and the testing MAE was improved to 1.01 ± 0.08%.
Conclusion: This is the first predictive model developed for ion chamber disagreements in patient-specific QA and was shown to be accurate within 3%. The model has been implemented in our clinical workflow and prospectively validated on 147 plans with MAE < 1%.
Funding Support, Disclosures, and Conflict of Interest: A.W. and G.V. report an ownership stake in Foretell Med, LLC, which is developing machine learning models in medicine
Handouts
Keywords
Quality Assurance, Validation, Modeling
Taxonomy
TH- External Beam- Photons: Quality Assurance - IMRT/VMAT
Contact Email