WE‐B‐BRA‐01: GPU‐Based Ultra Fast Direct Aperture Optimization in IMRT Treatment Planning

Purpose: To develop a graphics processing unit (GPU) based ultra fast direct aperture optimization (DAO) algorithm in intensity modulated radiation therapy (IMRT) treatment planning. Method and Materials: While DAO problems are well‐studied, most of them focus on heuristic search methods such as simulated annealing and the computational efficiency is often very low. In contrast, we formulate this problem as a large‐scale convex programming problem in terms of all multi‐leaf collimator (MLC) deliverable apertures and their associated intensities. An exact method called column generation method is used which iteratively adds high‐quality deliverable apertures while the treatment plan quality is improved. The efficiency of this approach has been ensured by exploiting fine‐grained parallelism to effectively utilize the computational recourses on GPU. Results: A clinical prostate IMRT case with various beamlet and voxel sizes is used to evaluate our implementation. High quality treatment plans are obtained at a high efficiency. In particular, for a 9‐field prostate case with 5×5 mm² beamlet size and 2.5×2.5×2.5 mm³ voxel size, the implementation takes only 7.1 seconds to generate 50 MLC apertures on an NVIDIA Tesla C1060 GPU. Conclusion: This work solved a major problem in developing ultra fast (re‐)planning technologies for fast treatment planning in conventional IMRT and for real‐time re‐planning for online adaptive radiotherapy.