近年來運用磁振造影技術來探討不同大腦功能區之間的複雜神經連結，成為神經科學研究、臨床神經醫學及精神疾病成因研究中相當重要的一個方法。擴散磁振影像(diffusion MRI)由於具有非侵入式偵測神經纖維組成、神經束髓鞘完整性及纖維走向的優點，更是廣泛地被應用於解析複雜的大腦神經連結。相較於確定型神經徑路追蹤演算法(deterministic tractography methods)，機率型神經徑路追蹤演算法(probabilistic tractography methods)的處理過程需要大量重複性的數學運算，因此難以快速呈現神經纖維追蹤結果並使用於臨床診斷應用。因此本研究之研究目的在於開發以圖形處理單元(Graphics Processing Unit，GPU)與計算統一架構(Compute Unified Device Architecture, CUDA)為基礎之機率型神經徑路追蹤演算法，以多核心平行運算縮短計算時間，提供快速呈現神經纖維連結機率與臨床應用之可行性。
本研究之研究成果顯示在進行大量隨機起始點之機率型神經徑路追蹤時，GPU平行運算相較於使用一般電腦之單一中央處理單元(CPU)，確實能夠大幅縮短計算時間，且於計算一千萬個起始點時，利用 NVidia Tesla C2075 圖形處理卡之448核心相較於使用單一核心Intel(R) Core CPU i5-2450M約可加速96倍。此外，為了驗證所開發之演算法的可信度與效能，本研究亦與英國牛津大學所發展之FMRIB Software Library (FSL) probtractx演算法進行神經追蹤機率結果之相關性分析與時間效能比較，結果顯示兩演算法所產出神經連結機率圖譜之相關性為0.6至0.7之間，而在十萬個隨機起始點之計算上， 本研究演算法於NVidia Tesla C2075 圖形處理卡之448核心相較於FSL probtractx演算法於IntelR XeonR Processor E5-2670 8核心CPU平行環境下約可加速3.5倍。
本研究成功開發了一個以GPU圖形運算單元為基礎的機率型神經徑路追蹤演算法，研究結果顯示相較於單一CPU與8 核心CPU之運算處理，本研究均能夠有效縮短計算處理時間，且演算法所所產出之神經連接機率圖譜與神經科學領域常用之FSL軟體之結果具有高度相關性，因此本演算法除了具有極佳之運算效能，且亦具有大腦神經纖維連結機率圖譜產出之可信度，相信未來將可廣泛地被運用於腦科學領域研究與更多之臨床應用。
In the past few years, Magnetic Resonance Imaging (MRI) techniques has become an important tool to explore the complicated structural and functional connectivity between different brain regions in the field of neuroscience, clinical neurology, neurosurgery and neuropsychiatry. Among these MRI technologies, diffusion MRI with tractography methods has been widely used to resolve complex neural tracts and delineate the neural pathways noninvasively due to its advantages with the abilities to measure the diffusion of water and provide the information of integrity, components and fiber orientation in neural tissues. In comparison with deterministic tractography methods, probabilistic tractography methods are more time-consuming because of the limitation of its massive computational requirements. Probabilistic tractography methods might be difficult to present the tracking results real-time and to be employed in clinical applications. Therefore, the purpose of this thesis is to develop a probabilistic tractography algorithm with Graphics Processing Unit (GPU) and Compute Unified Device Architecture (CUDA) platform by NVIDIA for accelerating the repeated mathematic calculations, extracting probabilistic connectivity mapping accurately and providing the feasibility of clinical applications by GPU parallel computing technique.
From the results, our algorithm using GPU parallel computing could show better performance in shortening the processing time comparing with single CPU sequential computing. In the assignment of ten millions seed points for fiber tracking, our algorithm with NVIDIA Tesla C2075 graphic card with 448 cores shows 96 times of accelerating performance compared to the same tractography method with Intel Core CPU i5-2450M. Moreover, in order to prove the reliability and to evaluate the performance of our algorithm, we also employed the FMRIB Software Library (FSL) probtractx method developed by the Oxford University in England as reference. The same region of interesting with a hundred thousand seed points has been assigned for tractography in our implemented algorithm with NVIDIA Tesla C2075 and FSL probtractx with IntelR XeonR Processor E5-2670 8 cores CPU respectively. The results showed that the similarity of probabilistic connectivity mapping between two methods was in the range of 0.6~0.7 and our GPU tracking method achieves a speedup factor of up to 3.5 times.
In this research, we successfully implemented a GPU-based probabilistic tractography algorithm and presented the better performance of acceleration, higher feasibility and reliability of our method by comparing with single CPU method and the existed software-FSL with multiple CPU cores respectively. Our implementation provides the potential indeed to apply probabilistic tractography for the studies of neuroscience and clinical applications reliable and faster.
