Room: AAPM ePoster Library
Purpose: To demonstrate the utility of an artificial neural network (ANN) to predict MLC and gantry position errors occurring during VMAT delivery. Sculpted dose distributions characteristic of VMAT often require extensive fluence modulation and, thus, precise control over machine parameters is essential for accurate treatment delivery. Predictive models may facilitate the generation of treatment plans robust to errors in gantry and MLC leaf positions.
Methods: Errors in gantry and MLC leaf position were predicted using two separate ANN models: a gantry model composed of five hidden layers and a MLC model of three; both models utilizing 100 nodes/layer. Gantry and MLC leaf positions were predicted based on features characterizing position and velocity of each. Data from 18 VMAT trajectory files representing various treatment sites were used to train and validate (90%/10%) the ANN model with 10 additional files serving as the test set. Model error was assessed by comparing predicted values with values recorded in trajectory files. This error was compared with the error resulting from differences between treatment plan-derived and machine-reported component positioning.
Results: Across the test set, the average root-mean-square error (RMSE) for predicting MLC positional errors was 0.006 mm (0.004-0.012 mm) using the ANN model compared to 0.032 mm (0.014-0.038 mm) using treatment planning values. The average RMSE for predicting gantry positional errors was 0.017° (0.008-0.026°) using the model compared to 0.048° (0.038-0.061°) with the plan.
Conclusion: MLC and gantry positions reported by trajectory files were closer to the positions predicted by our models than to those expected according to the plan. Predicting MLC and gantry positions allow for preemptive assessment of treatment plan deliverability. This approach may be used to develop robust treatment plans as well as to develop meaningful plan-specific and periodic machine quality assurance assessments.