Program Information
An Efficient Approach to Impose Hard Constraints in Radiation Therapy Inverse Treatment Planning
X Liu*, AH Belcher, R Wiersma, The University of Chicago, Chicago, IL
Presentations
SU-E-FS2-1 (Sunday, July 30, 2017) 1:00 PM - 1:55 PM Room: Four Seasons 2
Purpose: Quasi-Newton methods such as L-BFGS-B can solve most radiation treatment planning problems with simple box constraints efficiently, but they cannot solve several next generation treatment planning solutions that use hard constraint and mixed-norm objective functions such as MV+kV and total variation regularization based optimization. Using other convex optimizers such as CVX and Matlab quadprog often leads to a prohibitively long run times. In this work, we proposed an efficient approach to impose hard dose constraints based on a recently developed proximal operator graph solver (POGS).
Methods: Two key components of POGS are proximal operators of objective functions and the projection of a point to a graph. The tumor and OAR doses were included in the objective function and can be viewed as soft constraints. For these parts of tumor and OAR, hard dose constraints can be imposed by adding an indicator function term to the objective function, and the corresponding new proximal operator equals to the original value if the dose is inside the dose range and equals the bound values otherwise. Most importantly, this approach does not increase memory requirements, and the computation time in each iteration remains the same.
Results: The approach was tested by the Common Optimization for Radiation Therapy (CORT) dataset, including TG119, prostate and liver. All optimizations were solved by POGS, Matlab quadprog, CVX solvers, and problems without hard dose constraints were additionally solved by L-BFGS-B. The run time with and without hard constraints by POGS had no significant difference. POGS was about 30-60 times faster than other solvers for problems involving hard dose constraints.
Conclusion: An efficient approach to impose hard dose constraints for treatment planning was developed. It deals with hard constraints along with soft constraints in a way that does not increase memory requirements or computational times.
Contact Email: