Premium
SU‐GG‐T‐511: Towards Real‐Time Radiation Therapy: Superposition/Convolution at 4fps
Author(s) -
Jacques R,
Taylor R,
Wong J,
McNutt T
Publication year - 2008
Publication title -
medical physics
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 1.473
H-Index - 180
eISSN - 2473-4209
pISSN - 0094-2405
DOI - 10.1118/1.2962260
Subject(s) - computer science , computational science , kernel (algebra) , parallel computing , grid , cuda , algorithm , ray tracing (physics) , computer graphics (images) , physics , mathematics , optics , geometry , combinatorics
Purpose: To enable novel radiation therapy workflows by providing superposition/convolution dose calculation at interactive rates. Method and Materials: The Superposition/Convolution (S/C) dose computation algorithm utilizes ray‐tracing techniques extensively and is well suited to acceleration using consumer graphics processing units (GPU). The standard S/C algorithm was parallelized and implemented using NVIDIA's CUDA GPU programming environment. The basic S/C method generates incident fluence, using a source model, and transports it from the radiation source, using ray‐tracing, to deposit a TERMA grid. Then superposition spreads the dose deposition kernel out by ray‐tracing from each point in the TERMA grid. There are several issues with parallelizing this algorithm. Calculation of the TERMA grid suffers from read‐write conflicts. We solve this by running groups of spatially separated divergent rays. The forward S/C kernel also suffers from read‐write conflicts, but the inverse kernel does not. We reduced memory bandwidth by using a variable step ray‐tracer, independently attenuating each spectral energy bin and caching lookup‐tables. We explored the concept of using volumetric mip‐maps to approximate the ray as a true solid angle. Results: We compared our implementation to Pinnacle 3 for low (64×64×64 cube with 4 mm voxels) and high resolution (128×128×128 cube with 2mm voxels) water phantoms using as similar settings as possible. Pinnacle 3 (Philips ‐ Madison, WI) was run on a SunFire v250 machine with times of 30.961s and 365.944s respectively. We performed our calculations on a single NVIDIA GeForce 8800 GTX with times of 0.217s and 2.197s respectively. Preliminary experiments using volumetric mip‐maps showed additional performance improvements with minor accuracy loss. Conclusion: We have completed a GPU accelerated superposition/convolution dose engine, providing a substantial performance gain over CPU based implementations — indicating that real time dose computation is feasible with the accuracy levels of the S/C algorithm.