A deep learning method for predicting proton beam range and spread-out Bragg peak in passive scattering mode

It is difficult to calculate monitor units in the proton treatment planning system due to the complexity of using this system in the double scattering mode of proton therapy. Moreover, the range and spread-out Bragg peak (SOBP) values using the conversion algorithm (CONVALGO) provided by IBA (\({C}_{{\text{range}}}\), \({C}_{{\text{SOBP}}}\)) are different from the actual measured range (\({M}_{{\text{range}}})\) and SOBP (\({M}_{{\text{SOBP}}}\)) values. In this regard, the CONVALGO (FC) value (\({FC}_{{\text{range}}}\), \({FC}_{{\text{SOBP}}})\) should be measured according to the quality assurance (QA) of patient treatment, which requires physical effort and time. This study, therefore, aimed to reduce the time and effort spent on QA. The predictive model was trained using six parameters. Main option, sub-option, \({M}_{{\text{range}}}\) and \({M}_{{\text{SOBP}}}\) were used as input values, and \({FC}_{{\text{range}}}\) and \({FC}_{{\text{SOBP}}}\) were used as label. The trained model predicted the CONVALGO (PC) values of \({PC}_{{\text{range}}}\) and \({PC}_{{\text{SOBP}}}\). The test dataset has 261 patient data that were not used for training. Difference, mean absolute error (MAE), and root mean square error (RMSE) values were used for comparison. Compared to the FC value, the maximum difference was − 2.2 mm for \({PC}_{{\text{range}}}\) and − 3.4 mm for \({C}_{{\text{range}}}\). The acceptable standard of patient QA in our institute is within 1 mm and the number of data points that met the acceptable standard was 196 for \({PC}_{{\text{range}}}\) and 191 for \({C}_{{\text{range}}}\). For the MAE of \({PC}_{{\text{SOBP}}}\), options 1, 2, and 3 showed values within 1 mm. In the MAE of \({C}_{{\text{SOBP}}}\), the values were > 1 mm for all options.