Sampling issues for optimization in radiotherapy

A wide variety of optimization problems and techniques are used in radiation treatment planning. The problems typically involve large amounts of data, derived from simulations of patient anatomy and the properties of the delivery device. We investigate a three phase approach for their solution based on sampling of the underlying data that determines optimal beam angles, wedge orientations and delivery intensities in patient examples. Phase I uses multiple coarse samplings of the data and linear programming to adapt the sampling and determine a collection of promising angles to use. Phase II solves the adapted sample problems as mixed integer programs using only the promising angles. Phase III refines the sampling further, and fixes most of the discrete decision variables to reduce computation times. Particular emphasis will be given to general principles that are applicable to large classes of treatment planning problems. Specific examples show enormous increase in speed of planning, without detriment to the solution quality.