Accelerated evaluation of the robustness of treatment plans against geometric uncertainties by Gaussian processes
In order to provide a consistently high quality treatment, it is of great interest to assess the robustness of a treatment plan under the influence of geometric uncertainties. One possible method to implement this is to run treatment simulations for all scenarios that may arise from these uncertainties. These simulations may be evaluated in terms of the statistical distribution of the outcomes (as given by various dosimetric quality metrics) or statistical moments thereof, e.g. mean and/or variance. This paper introduces a method to compute the outcome distribution and all associated values of interest in a very efficient manner. This is accomplished by substituting the original patient model with a surrogate provided by a machine learning algorithm. This Gaussian process (GP) is trained to mimic the behavior of the patient model based on only very few samples. Once trained, the GP surrogate takes the place of the patient model in all subsequent calculations.The approach is demonstrated on two examples. The achieved computational speedup is more than one order of magnitude.