Room: AAPM ePoster Library
Purpose: To propose a planar dose calculation model for patient-specific IMRT QA using a simple convolutional neural network (CNN).
Methods: The CNN dose prediction model consists of four layers: input layer, convolution layer, batch normalization layer and regression layer. For each IMRT field, the input to the model is the weighted leaf map calculated using the multi-leaf collimator (MLC) leaf positions and fractional monitor units (MU) of individual segments; the output is the predicted planar dose map. The model was trained with reference dose maps calculated with an in-house, independent planar dose calculation software (MapCalc). Fifty-four IMRT fields were used to train the model; another thirty-four IMRT fields were used to evaluate the model’s accuracy. The predicted planar dose maps were compared with the planar dose maps calculated with both MapCalc and a commercial treatment planning system (TPS, Pinnacle3) using Gamma Passing Rate (GPR).
Results: The dose prediction model was successfully trained. The average GPR between the prediction and reference dose maps was 99.2±1.5% and 99.8±0.5% using 2%/2mm and 3%/3mm, respectively, for the training data, and 99.3±1.2% and 99.8±0.5% for the test data. The average GPR between the prediction and TPS dose maps was 95.8±3.3% (2%/2mm) and 98.5±1.6% (3%/3mm) for the training data, and 96.0±3.5% and 98.3±2.0% for the test data. These results showed that the prediction model generalized well on unseen IMRT fields. The training completed in 28 minutes using a personal computer equipped with a GPU and the prediction for a single IMRT field took 11 seconds on average.
Conclusion: We proposed a simple CNN dose prediction model that can accurately predict the planar dose maps for IMRT fields. The model, trained independently of the TPS, can be used to replace time-consuming measurement-based IMRT QA.
Quality Assurance, Convolution, Intensity Modulation