LAUSR

Purpose or Objective The COVID-19 pandemic forced radiation oncology departments to alter clinical workflows to reduce exposure risks in the clinic. Performing patient-specific quality assurance (PSQA) is one of the most resource intensive and time-consuming tasks. With technological advancements in radiotherapy treatment planning and quality assurance, research towards measurement-free PSQA has become a focus within the field. Most of these techniques involve modeling the relationship between treatment plan complexity and corresponding PSQA outcomes. However, to our knowledge, none of these efforts have been assessed and prospectively validated for clinical use. We implemented and a machine learning-based virtual VMAT QA (VQA) workflow to assess the safety and workload reduction of measurement-free patient-specific QA at a multi-site institution in light of COVID-19. Materials and Methods An XGBoost machine learning model was trained and tuned to predict QA outcomes of VMAT plans, represented as percent differences between the planned dose and measured ion chamber point dose in a phantom. The model was developed using a dataset of 579 previous clinical VMAT plans and associated QA measurements from our institution. 30 classes of complexity features were extracted from each VMAT plan and used as input for the model, which was tuned using a grid search over learning rate and tree depth hyperparameters and evaluated with 10-fold cross-validation. The final model was implemented within a webbased VQA application to predict QA outcomes of clinical plans within our existing clinical workflow. The application also displays relevant plan-specific feature importance and nearest neighbor analyses relative to database plans for physicist evaluation and documentation (Figure 1). VQA predictions were prospectively validated over one month of measurements at our clinic to assess the safety and efficiency gains of clinical implementation. $Φg Results 147 VMAT plans were measured at our institution over the course of one month, taking an average of approximately 20 minutes per plan for QA. VQA predictions for these plans had a mean absolute error of 0.97 +/- 0.69%, with a maximum absolute error of 2.75% (Figure 2). Employing a prediction decision threshold of 1% - meaning plans with absolute predictions of less than 1% would not need measurements - would flag all plans that may have ion chamber disagreements greater than 4%. This translates to a 73% reduction in QA workload in terms of time. A more conservative implementation of this workflow, where all SBRT plans will continue to be measured, would still result in a 46% reduction in QA workload. $Φg Conclusion To our knowledge, this is the first prospective clinical implementation and validation of VQA, which we observed to be safe and efficient. Using a conservative threshold, VQA can substantially reduce the number of required measurements for patient-specific QA, leading to a more effective allocation of clinical resources.