Implementation of Automatic Differentiation Tools for Multicriteria IMRT Optimization

Abstract

Automatic differentiation tools (ADOL-C) have been implemented for large-scale NLP optimization problems encountered in an advanced radiotherapy technique called Intensity Modulated Radiation Therapy (IMRT). Since IMRT treatments involve many tissue structures and their associated clinical objectives, the corresponding optimization problems are typically multi-objective. In this study, they are solved by a multi-criteria approach called Lexicographic Ordering. This approach allows clinical objectives to be categorized into several priorities or levels, and optimization is performed sequentially in order of priority while keeping the previously optimized results constrained. As a result, the feasible solution region is gradually reduced as the method progresses. For each level of optimization, the objective function and constraints are constructed interactively by a treatment planner and the corresponding Jacobian is provided by AD tools at a machine-precision level. Results indicate that a high degree of accuracy for Jacobian is essential to produce both feasible and optimal results for clinical IMRT optimization problems.