Click here to


Are you sure ?

Yes, do it No, cancel

Fast Direct Aperture Optimization Via Parallelizable Approximate Projection Onto the Set of Aperture-Ready Fluence Maps

D O'Connor1*, D Nguyen2 , D Ruan1 , A Landers1 , K Woods1 , E Boehnke1 , K Sheng1 , (1) UCLA, Los Angeles, CA, (2) UT Southwestern Medical Center, Dallas, TX


(Wednesday, 8/1/2018) 10:00 AM - 10:30 AM

Room: Exhibit Hall | Forum 9

Purpose: We present a direct aperture optimization (DAO) algorithm based on an efficient parallelizable method for computing, to a good approximation, the projection of a fluence map onto the collection of "aperture-ready" fluence maps.

Methods: The DAO problem is formulated as a standard fluence map optimization (FMO) problem with N fluence maps per beam, and an additional hard constraint that each fluence map must be "aperture-ready". A fluence map is "aperture-ready" if all non-zero intensity values are equal and, moreover, in each row there are no gaps between non-zero values. The significance of a fluence map being "aperture-ready" is that it can be delivered using a single multi-leaf collimator (MLC) aperture shape. The non-convex optimization problem is solved using an accelerated projected gradient algorithm. (Specifically, we use the fast iterative shrinkage-thresholding algorithm.) Our approach is based on the observation that projecting onto the set of aperture-ready fluence maps would be easy if the non-zero intensity value were known in advance. Thus, we can compute the projection approximately by doing a brute-force search over the range of possible intensity values for the fluence map. This approximate projection step is highly parallelizable, as it separates into independent subproblems for each fluence map, and moreover each candidate intensity value can be evaluated by its own worker.

Results: The algorithm was tested by creating coplanar 12-beam treatment plans for a C-shaped phantom with 5, 10, and 20 aperture shapes per beam. The optimization runtimes were 2.1, 2.7, and 4.3 minutes (respectively). We observed that performing the projection step in parallel with 12 threads decreased the runtime for the projection step by a factor of approximately 10. Further speedup is possible with more cores.

Conclusion: We have presented a promising approach to direct aperture optimization which makes full use of modern distributed computing platforms.

Funding Support, Disclosures, and Conflict of Interest: NIH U19AI067769 DE-SC0017687 NIH R21CA228160 DE-SC0017057 NIH R44CA183390 NIH R43CA183390 NIH R01CA188300


Not Applicable / None Entered.


Not Applicable / None Entered.

Contact Email