Ultra-fast fluence optimization for beam angle selection algorithms

17th International Conference on the Use of Computers in Radiation Therapy

物理学;计算机科学

Bangert, M.^1 ; Ziegenhein, P.^1 ; Oelfke, U.^1

German Cancer Research Center, DKFZ Im Neuenheimer Feld 280, 69120 Heidelberg, Germany^1

关键词: Beam angle selection;    Beam orientation;    Clinical evaluation;    Computational task;    Gradient calculations;    Memory locations;    Multiprocessor workstations;    Objective functions;   

Others  :  https://iopscience.iop.org/article/10.1088/1742-6596/489/1/012044/pdf
DOI  :  10.1088/1742-6596/489/1/012044

学科分类：计算机科学（综合）

来源: IOP

【 摘 要 】

Beam angle selection (BAS) including fluence optimization (FO) is among the most extensive computational tasks in radiotherapy. Precomputed dose influence data (DID) of all considered beam orientations (up to 100 GB for complex cases) has to be handled in the main memory and repeated FOs are required for different beam ensembles. In this paper, the authors describe concepts accelerating FO for BAS algorithms using off-the-shelf multiprocessor workstations. The FO runtime is not dominated by the arithmetic load of the CPUs but by the transportation of DID from the RAM to the CPUs. On multiprocessor workstations, however, the speed of data transportation from the main memory to the CPUs is non-uniform across the RAM; every CPU has a dedicated memory location (node) with minimum access time. We apply a thread node binding strategy to ensure that CPUs only access DID from their preferred node. Ideal load balancing for arbitrary beam ensembles is guaranteed by distributing the DID of every candidate beam equally to all nodes. Furthermore we use a custom sorting scheme of the DID to minimize the overall data transportation. The framework is implemented on an AMD Opteron workstation. One FO iteration comprising dose, objective function, and gradient calculation takes between 0.010 s (9 beams, skull, 0.23 GB DID) and 0.070 s (9 beams, abdomen, 1.50 GB DID). Our overall FO time is