基於異質運算架構的高效能導管式光學同調斷層掃描術

Abstract

本論文以導管式光學同調斷層掃描術(catheter-based Optical Coherence Tomography, OCT)為應用核心，設計並實現一套高效能異質運算系統。系統以AlazarTech ATS9373高速數位器為資料擷取平台，結合內建現場可程式化邏輯閘陣列(Field Programmable Gate Array, FPGA)即時進行快速傅立葉轉換(Fast Fourier Transform, FFT)，整合我們實驗室研究團隊所發展的1.3.2 圖形處理器(Graphics Processing Units, GPU)為主體之非均勻旋轉畸變(Non-Uniform Rotational Distortion, NURD)修正及光學同調斷層血管造影術(OCT Angiography, OCTA)演算法處理流程，並於CPU運算過程中導入OpenMP平行化技術，協同提升影像前處理效率。 本系統將資料擷取、即時FFT運算、NURD修正、OCTA分析及顯示前的影像加速等流程，分別由FPGA、GPU及CPU協同負責，完整串接即時catheter-based OCT影像資料流，展現異質運算平台於高效醫學影像處理之優勢。 論文內容涵蓋OCT原理、導管式影像NURD修正與OCTA演算法設計，並詳述系統架構、軟體整合流程、影像優化方法。針對FPGA-based FFT於即時成像應用中所遇挑戰，提出包括CPU/OpenMP加速、GPU-NURD修正及GPU-OCTA整合等系統架構與解決方案，進一步提升影像品質與使用者介面(GUI)操作便利性。實驗結果證明，所建構系統於即時成像、影像修正及介面操作等層面均有效提升效能，並增進影像品質及血管顯示效果。

This thesis focuses on catheter-based Optical Coherence Tomography (OCT) and presents the design and implementation of a high-performance heterogeneous computing system. The system employs an AlazarTech ATS9373 high-speed digitizer as its data-acquisition platform, featuring a built-in Field-Programmable Gate Array (FPGA) for real-time Fast Fourier Transform (FFT) processing. In addition, GPU-based algorithms for Non-Uniform Rotational Distortion (NURD) correction and Optical Coherence Tomography Angiography (OCTA) developed by our research group are integrated, with Graphics Processing Units (GPUs) serving as the primary computational engines. OpenMP parallelization is employed on the CPU to further enhance image-preprocessing efficiency. The proposed architecture assigns data acquisition, real-time FFT computation, NURD correction, OCTA analysis, and pre-display image acceleration to the FPGA, GPU, and CPU, respectively, thereby enabling a seamless real-time catheter-based OCT dataflow. This design highlights the advantages of heterogeneous computing platforms for high-efficiency medical image processing. The thesis covers OCT fundamentals, the design of NURD-correction and OCTA algorithms for catheter-based imaging, and provides detailed descriptions of the system architecture, software-integration workflow, and image-optimization strategies. To address the challenges of FPGA-based FFT in real-time imaging, a system architecture and complementary solutions, including CPU/OpenMP acceleration, GPU-implemented NURD correction, and GPU-OCTA integration, are proposed to further improve image quality and enhance graphical user-interface (GUI) usability. Experimental results demonstrate that the developed system significantly improves performance in real-time imaging, image correction, and interface operation, while delivering higher-quality images and clearer vascular visualization.