Machine Learning for Predicting Dose Inaccuracies in RapidArc Plans

The study addresses discrepancies between measured and calculated doses in RapidArc plans, especially in cases with small PTV or overmodulated MLC patterns. The goal was to create a Machine Learning (ML) model predicting dose inaccuracies in plans affected by overmodulation.
70 stereotactic RapidArc plans were planned in Eclipse (v.13.7), Delivery Quality Assurance (DQA) was performed, and doses were measured using PTW SRS1000 chamber array. Features from RPDicom files capturing plan complexity were used to train a Gaussian process regression model with 5-fold cross-validation and PCA. Open-loop validation on 25 plans not in the training set showed strong performance (RMSE 0.003, R2 0.96).
The results obtained indicate both a good accuracy and good generality of the trained model. This method shows promise in predicting plan deliverability, potentially reducing DQA failures, re-optimizations, and plan rejections due to overmodulation issues during physics plan checks. It proves to be a valuable tool in streamlining and enhancing the reliability of radiotherapy planning.