Portal dose image prediction using Monte Carlo generated transmission energy fluence maps of dynamic radiotherapy treatment plans: A deep learning approach

<i>Aims.</i>This work aims to develop and investigate the feasibility of a hybrid model combining Monte Carlo (MC) simulations and deep learning (DL) to predict electronic portal imaging device (EPID) images based on MC-generated exit phase space energy fluence maps from dynamic radiotherapy treatment plans. Such predicted images can be used as reference images during<i>in vivo</i>dosimetry.<i>Materials and methods</i>. MC simulations involving a Varian True Beam linear accelerator model were performed using the EGSnrc code package. Two custom variants of the U-Net architecture were employed. The MLC dynamic chair sequence and 17 clinical treatment plans, spanning various cancer types and delivery methods, were used to acquire experimental data, and in the MC simulations. The proposed method was tested through 2D gamma index analysis, comparing predicted and measured EPID images.<i>Results</i>. Results showed gamma passing rates of 38.65%, 74.16% and 96.17% (minimum, median, maximum) for a simpler model variant and 52.72%, 80.61% and 96.80% for the more complex model variant.<i>Conclusion</i>. The study highlights the feasibility of integrating MC and DL methodologies for<i>in vivo</i>dosimetry quality assurance in complex radiotherapy delivery.