Background: Accurate and rapid surface dose calculation is critical in superficiallow-energy electron radiotherapy due to the shallow treatment depth and the riskof radiation-induced skin toxicity. Traditional Monte Carlo (MC) simulations,while precise, can be computationally expensive and time-consuming.
Methods: To improve both the accuracy and speed of dose calculations, thisstudy combined MC simulations with deep learning techniques. Low-energy electron beams were simulated for six body sites using DOSXYZnrc, producing CTphantoms and corresponding dose distributions. A cascaded 3D-UNet (C3D) model was trained on these MC simulation datasets to predict dose distributionsrapidly.
Results: The C3D model demonstrated a mean absolute percentage error(MAPE) of less than 8% for one-dimensional dose curves, with no significantdifferences found in t-tests. The maximum dose difference observed across different body sites in 2D slices was 0.0421 ± 0.0340. The model achieved an overallGamma pass rate of over 92.09 ± 0.51% within 1%/1mm tolerance, and fordose distribution (DD) analysis under the 1% tolerance, the pass rate was 93.58± 0.21%. Additionally, the C3D model completed dose predictions in just 0.42seconds, making it approximately 140,000 times faster than MC simulations.
Conclusion: The integration of deep learning with MC simulations significantlyenhances the efficiency of surface dose calculations in superficial electron radiotherapy. The C3D model provides rapid and accurate dose predictions, facilitatingmore efficient treatment planning while maintaining high accuracy compared totraditional MC methods.