利用圖形處理器加速基於蒙地卡羅演算法之輻射劑量計算模擬之研究 - 從巨觀劑量到微觀劑量

Abstract

輻射劑量學是一門研究如何測量以及計算輻射劑量的科學，在醫學物理的研究上佔有相當重要的地位。輻射劑量的研究主要可分為兩個領域 – 巨觀劑量計算及微觀劑量計算。巨量劑量法與吸收劑量有關，吸收劑量率定義為忽略所得隨機波動的統計平均值，廣泛用於癌症放射治療中以模擬癌症治療所需的信息，例如輻射強度，照射時間和入射方向。微觀劑量法是對被輻照物質內之微觀結構中吸收能量的時空分佈下所作的系統研究，例如通過電離輻射產生的自由基分佈到人體內，進而破壞DNA並導致癌症的形成。 蒙地卡羅法為現今公認最為準確的模擬方法，此法採用一步一步計算粒子的行進過程進行計算。為了取得所希望的統計準確性，必須一次模擬大量粒子，這使得劑量計算需涉及模擬大量粒子行進及物理、化學作用，而耗費非常長的計算時間。 為了有效率的計算，往往需要建置龐大電腦運算資源。今日各醫學臨床及研究機構多使用叢集電腦實現平行劑量計算以減少運算時間。但一般研究人員要取得叢集電腦系統並不容易，因此阻礙了許多先進的輻射劑量學相關研究。然而近年來GPU的出現，對於平行運算提供了另一種高效能且成本低廉的選擇。 在本論文中，通過GPU的並行計算，成功實現了電子，光子及質子等多種放射治療劑量計算的加速。我們也成功結合商用軟體與碳離子劑量計算模擬器，期望為致力於基於GPU架構進行劑量學研究的研究人員和臨床醫生提供一個友好界面。對於微觀劑量法，我們開發出 gMicroMC，這是一種基於GPU架構的水輻解微觀蒙地卡羅模擬工具。據我們了解，這也是首度於GPU架構上，實現水輻解化學階段的蒙地卡羅模擬器。我們最終完成的GPU架構gMicroMC模擬器，完整實現物理，物理化學和化學三個階段，並進一步對其進行性能評估研究，得到整體模擬效能可獲得高於千倍於單一CPU執行緒的加速成果。 為了使整個開發過程更具通用性，我們提出了基於蒙地卡羅和GPU加速的輻射劑量學聯合程式設計模型(Joint programming model)。藉由歸納不同粒子類型的劑量計算研究，將其整理到一個通用的實現流程中，通過分析每個蒙地卡羅步驟中的行為，最後提出了有利於GPU架構的最佳化實現方法。

Radiation dosimetry is a science that studies how to measure and calculate radiation doses. It plays an important role in the research of medical physics. Existing MC dosimetry methods for radiation therapy could be broadly divided into two domains of macrodosimetry and microdosimetry. Macrodosimetry is concerned with absorbed dose and absorbed dose rate are defined as statistical averages that disregard the resulting random fluctuations and widely used in cancer radiotherapy to simulate the information needed for cancer treatment, such as radiation intensity, exposure time, and incident direction. Microdosimetry regards the systematic study of the spatial and temporal distribution of the absorbed energy in microscopic structures within the irradiated matter, such as the distribution of free radicals generated by ionizing radiation into the human body, damaging DNA and leading to the formation of cancer. Monte Carlo method is currently recognized as the most accurate simulation method. This method uses a step-by-step calculation of the particle travel process. In order to obtain the desired statistical accuracy, a large number of particles need to be simulated at one time. This makes the dose calculation involve a large number of particle travel and simulation of physical and chemical effects, which consumes very long calculation time. Huge computer computing resources need to be utilized to perform efficient Monte Carlo calculations. Today, various medical clinical and research institutions use cluster computers to achieve parallel dose calculations to reduce computing time. However, it is not easy for general researchers to build a cluster computer system, which hinders advanced radiation dosimetry research. The emergence of GPUs in recent years has provided another high-performance and low-cost option for parallel computing. In this dissertation, we have successfully achieved the acceleration of various radiation therapy dose calculations, for example, electron, photon and proton through the parallel computing of GPU. We have also successfully combined commercial software with a carbon ion dose calculation simulator, expecting to make a friendly interface for researchers and clinicians devoted to GPU-based dosimetry research. For the microdosimetry, gMicroMC, a GPU‐based microscopic Monte Carlo simulation tool for water radiolysis has been developed which is the first Monte Carlo method based on GPU to implement the water radiolysis chemical stage simulator to our knowledge. We finally realized all three stages-physics, physical chemistry and chemical stages based on the complete GPU-based gMicroMC simulator. The experimental results of evaluation on gMicroMC show up to three orders of magnitude performance gain. With the aim of making the entire development process more versatile, a joint programming model based on Monte Carlo with GPU acceleration for radiation dosimetry is proposed. We organized our studies on the dose calculation of different particle types into a generic implementation flow, optimized the performances by analyzing the behaviors in each Monte Carlo step, we present implementation methods that are conducive to the GPU architecture.