An Oracle Solution for Performance Benchmarking of Dynamic Multi-Leaf Collimator Algorithms
J Sherman*, P Keall,SU-E-T-329 Sunday 3:00PM - 6:00PM Room: Exhibit Hall
Purpose: Radiotherapy for tumors in motion is an area of active interest in Radiation Physics. To date most approaches develop a treatment plan with static information and then, using realtime tumor information attempt to translate the plan to keep the treatment synced with the tumor. If the tumor motion results in infeasible DMLC leaf trajectories the treatment fidelity is reduced. By performing retrospective analysis we are able to calculate the optimal DMLC leaf trajectories and use the optimal to derive improvements to the predictive methods as well as providing simply providing a means of measuring the realistic best case performance a predictive method can achieve.
Methods: For the analytical models the Euler-Lagrange method was used to solve for the optimal leaf trajectories incorporating physical constraints. The analysis of real patient data was performed by recasting the problem as a mixed integer programming problem implemented using C++ and CPLEX.
Results: From the theoretical analysis we found that when the leaf velocity constraint is violated the optimal leaf path is piecewise linear with the slope of the line defined by the leaf velocity constraint. From this theoretical work we were then able to show that anticipating the motion provides a 50% improvement in treatment fidelity. The numerical work provides a crucial best possible treatment, aka an oracle solution, and is a key component of future improvements in adaptive radiotherapy.
Conclusions: This work provides some initial simple rules for improving adaptive radiotherapy which will have impact on treatment fidelity. In addition we provide a method that retrospectively calculates the best possible leaf trajectories. This is of first order significance in improving treatments by learning from past data and allowing us to better understand the ways in which the current state of the art can be improved.