Predicting Measurement-Based Patient-Specific Quality Assurance Results for ViewRay Plans Using Machine Learning
A Witztum1*, G Ibrahim2, P Wall1, J Lamb2, G Valdes1, (1) University of California, San Francisco, San Francisco, CA, (2) University of California, Los Angeles, Los Angeles, CA
Presentations
SU-E-BRA-5 (Sunday, 7/10/2022) 1:00 PM - 2:00 PM [Eastern Time (GMT-4)]
Ballroom A
Purpose: Online adaptive radiotherapy systems, such as the ViewRay (ViewRay, Cleveland, Ohio), can create adapted treatment plans that are highly specific to daily anatomy and positioning. However, traditional pre-treatment measurement-based patient-specific quality assurance (PSQA) cannot be delivered to a phantom while the patient remains on the treatment couch. This work develops a machine learning model for predicting diode array gamma passing rates (GPRs) and ion chamber (IC) disagreements during PSQA of ViewRay treatment plans without the need for beam delivery.
Methods: A dataset of 58 ViewRay treatment plans and associated PSQA outcomes was retrospectively collected at a single institution. PSQA measurements were performed using an ArcCHECK (Sun Nuclear, Melbourne, FL) phantom with an inserted ion chamber at its center. Complexity features representing different failure modes were extracted from the upper and lower MLC banks from each DICOM RT-Plan file using in-house Python software (totaling 405 raw features). Two separate models were built to predict both GPRs and IC disagreements. In each case, a linear model with lasso regularization was trained and hyperparameters were tuned using 10-fold cross-validation with the glmnet MATLAB (MathWorks, Natick, MA) package.
Results: The R-squared values of the cross-validated test results were 0.38 for GPR and 0.33 for IC disagreement prediction models. The learning curves indicated that saturation had not been reached and the addition of more data should improve both models.
Conclusion: These are the first predictive models developed for an adaptive radiotherapy platform utilizing a double-focused MLC system, which show correlations between plan complexity features and both GPRs and IC disagreement from PSQA. This is the first demonstration of the feasibility of machine learning QA prediction in the online adaptive patient workflow. A larger multi-institutional dataset is being curated to improve accuracy and generalizability of these models.
Funding Support, Disclosures, and Conflict of Interest: AW and GV report an ownership stake of Foretell Med LLC who build machine learning models for PSQA.
Keywords
Taxonomy
TH- External Beam- Photons: adaptive therapy
Contact Email