Predictive Model of Difference Between Conventional and Monte Carlo Algorithms for Stereotactic Body Radiation Therapy of Lung Cancer

Abstract

To compare the conventional Ray Tracing (RT) and Monte Carlo (MC) dose calculations in patients treated with CyberKnife stereotactic ablative radiotherapy and find predictive model of RT/MC calculation difference prior to MC calculation. This is a study on 75 patients treated with stereotactic body radiotherapy of lung cancer. All plans were calculated with conventional RT algorithm and recalculated with MC algorithm using 1% uncertainty. The final MC plan was prescribed to the same coverage and same dose as previous RT plan (D95%). Two inhomogeneity boundaries (encloses the lung tissue - Lung shell, and the target - Target shell) were delineated as a shell structures ± 1 mm in/outside from the lung and ± 1 mm in/outside from the PTV. Lung shell is also proportional to size of lung. Volumes and mean doses of shell structures, numbers of monitor units per fraction (MU/fr), volumes of PTV and GTV, lung mean doses, volumes of lung tissue within the planning target volumes, tumor's localizations and differences in prescribed isodose lines between RT and MC were recorded. Doses and MU/fr were derived from plans calculated with conventional algorithm (RT). Statistical methods were used to find suitable variables (statistically significant) and for multivariable regression model delination. Cross-validation analysis was used to verify the model accuracy. The final prescribed isodose line was on average 80% (range 76-83) and 69% (range 53-78) for RT and MC algorithm respectively what confirmed that dose calculated with conventional algorithm is overestimated. Using the backward selection, a regression model for difference of RT/MC prescribed isodose line was built. Four regression variables (Lung shell, MU/fr, mean doses in Lung shell and Target shell) proved to be significant with p-values < 0,01. The R squared value of the model was 79%. Dividing data in two groups (estimation group consisted of 50 observations, evaluation group consisted of 25 observations) a cross-validation analysis of the model was provided. The cross-validation mean absolute error (CV MAE) equals to 1.8 (±1 in relation to MC 1% calculation uncertainty) and mean error 0.9. No single parameter can predict the difference between conventional (RT) and MC algorithms. The statistical analysis showed four significant variables which can be used in multivariable model to predict the RT/MC difference (RT overestimation). The value of CV MAE showed that our model can highlight the degree of difference from RT results prior to MC calculation. This is our preliminary result and further investigation will follow.