Streaming architectures for computation in medical radiation physics

点击翻译按钮获取中文摘要

Radiation therapy historically has evolved in close symbiosis with the development of faster and more powerful computers; advanced delivery techniques such as intensity-modulated radiation therapy (IMRT) would not be possible without today's microprocessors. The quest for better radiation delivery techniques continues to this day, and commands increasingly more powerful machines to handle the calculations required by complex treatment planning algorithms. In order to keep up with the processing requirements of modern algorithms, treatment planning systems (TPS) must sometimes resort to clusters of computers to provide clinical output in reasonable time. This architecture, although efficient, requires substantial investments in time and expertise, and can hinder the deployment of better TPS for economical and logistical reasons. It therefore appears that an economical and powerful, yet standalone calculation platform is highly desirable in radiation therapy, and in medical physics in general. We propose here to develop such a platform, using innovative hardware material dedicated to massively parallel calculations such as Graphics Processing Units (GPUs). GPUs and similar stream processors such as the Cell Broadband Engine (CBE) have already been used successfully for general-purpose calculations in many scientific fields, achieving speed improvement factors of up to two orders of magnitude over traditional implementations on Central Processing Units (CPUs). We propose to implement medical physics algorithms on stream processors such as GPUs in order to significantly accelerate tasks such as image processing, optimization and dose calculations. This, in turn, will allow the integration of more complex, more accurate and more personalized algorithms in treatment planning systems. The development of such algorithms on stream processors will benefit not only the field of medical radiation physics, but any discipline where high-performance computing is desirable.