Purpose: The dose weighted linear energy transfer (LETd) is one of the parameters defining the response of the tissue in proton therapy. Estimating LETd requires treatment planning information and time-consuming Monte Carlo (MC) simulations. We expand our reported artificial neural networks (ANN) three-dimensional dose prediction model to include MC dose and LETd.
Methods: 27 patient subset from an institutional craniopharyngioma treatment protocol comprised the study. All patients received a prescribed dose of 54 CGE. MC dose and LETd for model training were calculated using the Tool for Particle Therapy Simulation. Predicted beam path masks were created by expanding existing clinical target volume contours and raytracing along the beam directions. Using MATLAB’s neural network toolbox, voxel features within the predicted beam path were used to train feed forward ANN models that predicted dose and LETd. ANNs were trained with several network training and activation functions with varying number of deep layers and neurons. The final model is the one with lowest root mean square error (RMSE).
Results: The final trained model used 3 layers and 25 neurons with RMSE of predicted dose and LETd 5.8 CGE and 0.5 MeV/mm/(g/cm3), respectively. The percentage of voxels within ± 5% of the MC calculated values were 70.8% and 50.2% within the CTV, 44.9% and 38.1% 0-5mm from the CTV and 13.5% and 28.9% 5-15mm from the CTV in dose and LETd, respectively. Model agreement trended worse near sharp dose gradients, suggesting spatial uncertainty in the predicted values.
Conclusion: We implemented an ANN algorithm to predict dose and LETd. The benefits of upfront dose prediction are well documented in the knowledge-based planning community. Supplementing the treatment planning process with an LETd prediction allows for careful consideration of treatment geometries to limit the potential biologic impact of high LETd regions in normal tissues.
Not Applicable / None Entered.