A predictive quality assurance model for patient-specific gamma passing rate of hyperarc-based stereotactic radiotherapy and radiosurgery of brain metastases.

OBJECTIVE

Measurement-based patient specific quality assurance (PSQA) is an increasingly debated topic among medical physicists. Developments like online adaptive radiotherapy and same-day stereotactic treatments limit the time to do measurement-based PSQA. Herein, we develop a predictive machine learning model to supplement PSQA by predicting the gamma passing rate (GPR) per stereotactic arc. This streamlines PSQA, providing planners the insight to replan potentially sub-optimal plans, to mitigate machine time inefficiencies.

METHODS

122 patients that had previously received HyperArc stereotactic radiosurgery/radiotherapy on a TrueBeam LINAC (Millenium 120 MLCs, 6MV-FFF) were used to generate a long short-term memory (LSTM) recurrent neural network to predict the GPR for a 2%/2 mm criteria. GPRs were discretized into three classes: Ideal (≥95%), Investigate [85%-95%), and Replan (<85%). In total, 468 VMAT arcs were used for this model with a class distribution of 370 (Ideal), 65 (Investigate), and 33 (Replan). To counteract the imbalanced data, the minority classes were over-sampled using synthetic minority over-sampling technique to generate a balanced dataset. The LSTM model was trained in Python with an 80-20 training-testing stratified split. Individual class sensitivity and specificity were recorded following a one versus all method. The final model was deployed clinically through Eclipse Scripting.

RESULTS

The model demonstrated the following (sensitivity, specificity) for the testing data: Ideal (78.4%, 87.2%), Investigate (75.7%, 89.9%), and Replan (93.2%, 96.6%). The primary focus of this model is to identify failing beams and allow the planner to address this prior to running the PSQA, as such the Replan class was the most important for evaluation. A sensitivity of 93.2% indicates that the model will identify 93.2% of HyperArc plans that need to be replanned with a very high certainty due to the 96.6% specificity.

CONCLUSIONS

The predictive GPR model developed within this research enables HyperArc planners to immediately assess the GPR for each stereotactic arc and preemptively replan potentially failing arcs to optimize the PSQA machine time.