High-Performance Computing in Medical Physics

医学物理中的高性能计算

• 批准号: RGPIN-2018-04588
• 负责人: Després, Philippe
• 金额: $ 4.17万
• 依托单位: Université Laval
• 依托单位国家: 加拿大
• 项目类别: Discovery Grants Program - Individual
• 财政年份: 2020
• 资助国家: 加拿大
• 起止时间: 2020-01-01 至 2021-12-31
• 项目状态: 已结题
• 来源: https://www.nserc-crsng.gc.ca/ase-oro/Details-Detailles_eng.asp?id=713152
• 关键词: Performance; Computing; Medical; Physics

Context Several numerical methods are used in Medical Physics, ranging from purely analytical approaches to Monte Carlo techniques, and dedicated to tasks such as dose calculations, tomographic reconstruction and image processing. In general, the accuracy and the quality of solutions obtained with these methods increases with the realism (and complexity) of the model used to describe the physics of the problem. However, the long execution times associated with physics-rich models are usually not compatible with clinical workflows. In order to deploy physics-rich, more accurate algorithms in clinical environments, my research group has developed High-Performance Computing (HPC) approaches based on massively parallel Graphics Processing Units (GPUs) for several computing tasks in Radiology, Radiation Oncology, and Nuclear Medicine. For the 2018-2023 Discovery cycle, we will innovate by integrating data-driven approaches in Medical Physics applications that can benefit from the latest hardware and software developements in data sciences. Objective The overall objective of the proposed research program is to integrate advanced physics models as well as data-driven knowledge and techniques into Medical Physics applications in order to obtain better, more accurate solutions (images, dose maps) in clinically acceptable timeframes. Scientific approach We will continue the development of advanced, GPU-based tomographic reconstruction algorithms in CT and CBCT, with the objective of improving images while reducing the patient dose through a better handling of sparse and noisy data, notably through scatter corrections made possible by our fast GPU-based Monte Carlo code. Promising approaches using Convolutional Neural Networks (CNNs) will be used for noise regularization, in complement to the Total Variation approach we have used so far. We also plan to use the latest generation of machine learning-based image segmentation algorithms to automatically generate the contours of anatomical structures in CT data sets. The training of this algorithm, implemented with the TensorFlow framework, will proceed with existing high-quality and large data sets of images contoured by specialists at our institution. We also plan to combine these contours with our GPU-based Monte Carlo dose calculation engine to compute doses to organs in CT, which is the fundamental requirement for large scale population studies investigating the risk of radiation exposure from medical sources. The proposed research program also comprises activities at the interface of Nuclear Medicine and Radiation Oncology, aimed at developing a robust dosimetry and treatment planning platform for personalized radionuclide targeted therapy. Significance This research program will ultimately benefit patients and stakeholders in the healthcare sector, as well as industrial partners.

背景在医学物理学中使用了几种数值方法，从纯粹的分析方法到蒙特卡罗技术，并致力于诸如剂量计算、层析重建和图像处理等任务。一般来说，用这些方法获得的解的准确性和质量随着用来描述问题的物理模型的真实性(和复杂性)而增加。然而，与物理丰富的模型相关的长执行时间通常与临床工作流不兼容。为了在临床环境中部署物理丰富、更精确的算法，我的研究小组开发了基于大规模并行图形处理单元(GPU)的高性能计算(HPC)方法，用于放射学、放射肿瘤学和核医学的几个计算任务。在2018-2023年的发现周期中，我们将通过在医学物理应用中整合数据驱动的方法进行创新，这些方法可以受益于数据科学的最新硬件和软件开发。 目的研究计划的总体目标是将先进的物理模型以及数据驱动的知识和技术集成到医学物理学应用中，以便在临床可接受的时间范围内获得更好、更准确的解决方案(图像、剂量图)。 科学方法我们将继续在CT和CBCT中开发先进的基于GPU的断层扫描重建算法，目标是通过更好地处理稀疏和噪声数据，特别是通过我们基于GPU的快速蒙特卡罗代码实现的散射校正，在改善图像的同时减少患者剂量。使用卷积神经网络(CNN)的有前途的方法将被用于噪声正则化，作为我们到目前为止使用的总变分方法的补充。我们还计划使用最新一代的基于机器学习的图像分割算法来自动生成CT数据集中解剖结构的轮廓。利用TensorFlow框架实施的这一算法的训练将使用我们机构专家绘制的现有高质量和大型图像数据集进行。我们还计划将这些等高线与我们基于GPU的蒙特卡罗剂量计算引擎相结合，以计算CT中器官的剂量，这是调查医疗源辐射暴露风险的大规模人口研究的基本要求。拟议的研究计划还包括核医学和放射肿瘤学接口的活动，旨在为个性化放射性核素靶向治疗开发一个强大的剂量测定和治疗计划平台。 意义这项研究计划最终将使患者和医疗保健部门的利益相关者以及行业合作伙伴受益。