臺灣博碩士論文加值系統

隨著錐形射束電腦斷層掃描（CBCT）在牙科和放射治療應用中變得越來越普遍， CBCT製造商不得不面臨開發成本問題。除了CBCT機器之外，還需要提供使用者高速且昂貴的工作站和重建軟體。雲端運算的出現可以提供處理此問題的方法。由於所有硬體和軟體都在雲端上，最新的硬體和軟體可以通過網路提供給用戶，然而系統程式卻不需要改變。對於重建軟體，如何減少CBCT系統的輻射劑量仍然是需要關注的主題。利用疊代式影像重建演算法，例如最大相似期望值最大化算法（MLEM）的有限角度重建已經可以在減少所需的投影角度下提供滿意的影像品質，也同時減少了輻射劑量。然而，疊代式重建演算法需要龐大的運算能力。為了應對這一挑戰，我們提出了一種使用雲端GPU的CBCT重建系統，用於測試和實現疊代式重建方法。方法: 我們已經為重建系統設置客戶端和雲端軟體。在客戶端，軟體設定為CBCT每次完成新的造影時自動將投影數據上傳到雲端。在雲端上，我們已經實現了作業管理，並分別為FDK和MLEM提供了用於影像重建的非疊代和疊代式演算法。兩者都以CUDA語言實作，使其可以使用圖形計算單元（GPU）進行加速。測試雲端運算系統配置在配備8張GPU （NVIDIA GeForce GTX 1080T）的伺服器上，並使用Google Cloud Platform（GCP）進行驗證。結果: 在全角度重建評估中，與使用單一CPU的工作站進行重建相比，單張GPU可以將FDK和MLEM分別加速1223和2938倍。在本地伺服器及GCP上使用多張GPU可以有更進一步的加速效果，不過受限於重建影像的大小，過多的GPU使用數量反而會降低效能。在有限角度的重建評估中，在60個投影角度和50次疊代下，MLEM可以獲得使用最少角度下的令人滿意的影像品質。此外，在使用計算能力接近的GPU (1080Ti 和 P100) 重建的速度上，雲端GPU與本地GPU伺服器的效能相仿。在不同GPU數量的組合上，均可以使MLEM的速度降到一分鐘以內。結論: 通過雲端運算，可以有效地降低高級伺服器和重建軟體的成本。使用CUDA撰寫而成的重建軟體具有顯著的加速。該加速度對於測試新穎的疊代影像重建是很有用的，並可提供未來將有限角度重建應用在臨床CBCT影像的可能性。使用雲端GPU的CBCT疊代重建演算法，可以使CBCT的開發和測試更為快速，並有利於在劑量減少的情況下，使用有限角度CBCT重建的實際臨床應用。
關鍵字: 錐狀射束電腦斷層、疊代式重建演算法、有限角度、CUDA、多GPU重建、 雲端運算

Cone-beam computed tomography (CBCT) has become more and more popular in dental and radiation therapy applications. The CBCT vendors face a problem with development cost. Except for the CBCT scanner itself, a high-end and costly workstation with reconstruction programs also needs to be provided to the user. Cloud computing could provide a solution for this problem. With all the hardware and software on the cloud, the newest hardware and software can be provided to the user through the Internet, and the local program does not need to change. For the reconstruction program, dose reduction for CBCT systems remains to be a topic of concern. Limited angle reconstruction with iterative image reconstruction algorithms, such as maximum likelihood expectation–maximization algorithm (MLEM), has shown promises by yielding a satisfactory image quality while reducing the required projection angles and hence the radiation dose. However, iterative reconstruction algorithms require much more computational power. To address this challenge, we propose a cloud GPU-based reconstruction system to implement iterative reconstruction methods in this work. Methods: We have implemented both the client- and cloud-side programs for the reconstruction system. On the client side, the software is configured to automatically upload the projection data to the cloud whenever a new CBCT acquisition is completed. On the cloud side, we have implemented a job manager to provide the non-iterative and iterative algorithms for image reconstruction with FDK and MLEM, respectively. Both algorithms have been implemented as CUDA programs to accelerate the reconstruction with graphical computing units (GPUs). The testing cloud-computing system was configured on a server equipped with 8 GPU cards (NVIDIA GeForce GTX 1080Ti), and we also used Google Cloud Platform (GCP) for verification. Results: In the full-angle evaluation, the single GPU has 1223 and 2938 times acceleration with FDK and MLEM respectively, compared to a standalone CPU-based workstation. Using multiple GPUs on the local server and GCP could further accelerate performance. Due to the limitation of reconstruction volume, the excessive usage number of GPUs would reduce the property. In the limited-angle reconstruction tests, MLEM with 60 projection angles and 50 iterations has a satisfied image quality under the lowest numbers of angles. While using GPU that has similar computing ability, the local server and GCP could have similar speeds. For the combination of the different number of GPUs, the speed of the MLEM can be reduced to less than one minute. Conclusion: With cloud computing, the cost of the high-end workstation and reconstruction software can be greatly reduced. The reconstruction program implemented with CUDA has a significant acceleration. This acceleration is useful for testing novel iterative image reconstruction, as well as making future application of limited-angle reconstruction of clinical CBCT images a routine possibility. With cloud GPU-based environments, CBCT iterative reconstruction methods can be developed and tested in a faster pace that may accelerate the actual clinical application of limited angle reconstruction for CBCT dose reduction.

Keywords: CBCT, iterative reconstruction, limited angle, CUDA, multiple GPU reconstruction, cloud computing

Chapter 1 Overview 1
Chapter 2 Introduction 3
2.1 Background of CBCT 3
2.2 CBCT image reconstruction 5
2.3 Limited angle reconstruction 7
2.4 Accelerate image reconstruction with a graphics processing unit (GPU) 9
2.5 Multi-GPU 12
2.6 CBCT reconstruction using cloud computing environment 13
2.7 Google Cloud Platform(GCP) 15
2.8 Specific aims 17
Chapter 3 Materials and Methods 19
3.1 GPU acceleration and cloud computing 19
3.1.1 Local cloud compute environment 20
A. Hardware 20
B. Cloud reconstruction architecture 22
3.1.2 Google Cloud Platform(GCP) 23
3.2 Reconstruction algorithms 25
3.2.1 FDK 25
3.2.2 Maximum likelihood expectation–maximization algorithm (MLEM) 28
3.3 Multi-GPU acceleration 30
3.4 Prototype CBCT system 30
3.4.1 X-ray tube 31
3.4.2 Detector 32
3.4.3 Rotating platform 34
Chapter 4 Results and Discussion 36
4.1 Cloud-Based CBCT Reconstruction System 36
4.1.1 Client 37
4.1.2 Job manager and worker 37
4.1.3 Reconstruction Algorithm Software 38
A. FDK 38
B. MLEM 40
4.1.4 Other setting 42
4.2 Experimental results 42
4.2.1 Full angles speed comparison between single GPU and CPU 43
4.2.2 Full angles speed comparison between different numbers of GPU 45
4.2.2.1 Using 8-GPU server 45
4.2.2.2 Using GCP 51
4.2.3 Image quality comparison between different iteration and angles 54
4.2.4 Limited angles speed comparison between different numbers of GPU 62
4.2.4.1 Using 8-GPU server 62
4.2.4.2 Using GCP 65
4.3 Summary 67
Chapter 5 Conclusion 69
References 70

[1]W. C.Scarfe andA. G.Farman, “What is Cone-Beam CT and How Does it Work?, Dent. Clin. North Am., vol. 52, no. 4, pp. 707–730, 2008.
[2]L. A.Feldkamp, L. C.Davis, andJ. W.Kress, “Practical cone-beam algorithm, Josa a, vol. 1, no. 6, pp. 612–619, 1984.
[3]M.Beister, D.Kolditz, andW. A.Kalender, “Iterative reconstruction methods in X-ray CT, Phys. Medica, vol. 28, no. 2, pp. 94–108, 2012.
[4]G.Li, “Patient radiation dose and protection from cone-beam computed tomography, Imaging Sci. Dent., vol. 43, no. 2, pp. 63–69, 2013.
[5]T. L.Slovis, “The ALARA concept in pediatric CT intelligent dose reduction, Pediatr. Radiol., vol. 32, no. 4, pp. 219–220, 2002.
[6]H.Matikka andT.Virén, “Radiation dose reduction in cone-beam computed tomography of extremities: Evaluation of a novel radiation shield, J. Radiol. Prot., vol. 34, no. 2, 2014.
[7]A.AlNajjar et al., “Comparison of adult and child radiation equivalent doses from 2 dental cone-beam computed tomography units, Am. J. Orthod. Dentofac. Orthop., vol. 143, no. 6, pp. 784–792, 2013.
[8]X.Han et al., “Algorithm-enabled low-dose micro-CT imaging, IEEE Trans. Med. Imaging, vol. 30, no. 3, pp. 606–620, 2011.
[9]X.Jia, P.Ziegenhein, andS. B.Jiang, “GPU-based high-performance computing for radiation therapy, Phys. Med. Biol., vol. 59, no. 4, 2014.
[10]M.Leeser, S.Mukherjee, andJ.Brock, “Fast reconstruction of 3D volumes from 2D CT projection data with GPUs, BMC Res. Notes, vol. 7, no. 1, pp. 1–8, 2014.
[11]G.Pratx andL.Xing, “GPU computing in medical physics: A review, Med. Phys., vol. 38, no. 5, pp. 2685–2697, 2011.
[12]D.Kirk, “NVIDIA cuda software and gpu parallel computing architecture, Proc. 6th Int. Symp. Mem. Manag. - ISMM ’07, pp. 103–104, 2007.
[13]NVIDIA, “CUDA Toolkit Documentation. [Online]. Available: https://docs.nvidia.com/cuda/.
[14]M.Harris, “How to Overlap Data Transfers in CUDA C/C++, 2012. [Online]. Available: https://devblogs.nvidia.com/how-overlap-data-transfers-cuda-cc/.
[15]G. C.Kagadis et al., “Cloud computing in medical imaging., Med. Phys., vol. 40, no. 7, p. 070901, 2013.
[16]B.Meng, G.Pratx, andL.Xing, “Ultrafast and scalable cone-beam CT reconstruction using MapReduce in a cloud computing environment, Med. Phys., vol. 38, no. 12, pp. 6603–6609, 2011.
[17]“Care Cloud. [Online]. Available: https://www.carecloud.com/.
[18]“Oracle Healthcare Cloud. [Online]. Available: https://www.oracle.com/industries/healthcare/index.html.
[19]“Artificial Intelligence for Medicine ＆ Health Innovations. [Online]. Available: https://aimhi.tw/.
[20]Google, “Google Cloud Platform Documentation. [Online]. Available: https://cloud.google.com/docs/.
[21]L. A.Shepp andY.Vardi, “Maximum Likelihood Reconstruction for Emission Tomography, IEEE Trans. Med. Imaging, vol. 1, no. 2, pp. 113–122, 1982.
[22]“NVIDIA. [Online]. Available: https://www.nvidia.com/zh-tw/.
[23]B.Philippa, “matlab-job-manager. [Online]. Available: https://github.com/bronsonp/matlab-job-manager.
[24]O.Iakushkin andV.Grishkin, “Messaging middleware for cloud applications: Extending brokerless approach, 2014 2nd Int. Conf. Emiss. Electron. ICEE 2014 Joined with 10th Int. Vac. Electron Sources Conf. IVESC 2014, Int. Conf. Comput. Technol. Phys. Eng. Appl. ICCTPEA 2014, 20th Int. Work. Beam Dyn. Optim. BDO 2014 - Proc., pp. 4–7, 2014.
[25]Im.Corporation, “ZeroMQ. [Online]. Available: http://zeromq.org/.
[26]G.Sleijpen, “Universiteit Utrecht Master Thesis for Scientific Computing Image reconstruction methods for emission based tomography, 2011.
[27]M. N. and J. N. A.Wernick, Emission Tomography : The Fundamentals of PET and SPECT. 1900.
[28]P. B.Noël, A. M.Walczak, J.Xu, J. J.Corso, K. R.Hoffmann, andS.Schafer, “GPU-based cone beam computed tomography, Comput. Methods Programs Biomed., vol. 98, no. 3, pp. 271–277, 2010.
[29]X.Jia, B.Dong, Y.Lou, H.Kim, J.Chen, andA.Wang, “TIGRE : a MATLAB-GPU toolbox for CBCT image reconstruction, 1976.