利用圖形處理器加速引擎於高速導管式光學同調斷層掃描術之開發

Abstract

導管式光學同調斷層掃描術(Catheter-based optical coherence tomography, catheter-based OCT)的發展結合了OCT所具備的優勢以及具有可以被使用在人體內部的設備條件，使其有了作為對於人體內部器官前期病變早期診斷工具的可行性。然而，由於catheter-based OCT本身透過光點掃描成像機制的影響，會使得擷取到的影像出現畸變。此外，影像畸變的現象會導致進行進階的影像分析步驟時，可能會出現不精確的運算結果，使得catheter-based OCT要作為診斷的工具會具有一定的挑戰性。 在本篇論文中，我們使用圖形處理器(Graphics Processing Units, GPU)進行catheter-based OCT影像修正演算法的開發。影像修正的部分分為兩項，第一項是針對如光纖彎曲等原因造成照射在樣本表面之入射光偏振態變化所產生的偽影進行修正，除了使用偏振分集檢測模組蒐集來自兩種偏振態的干涉訊號外，我們還改寫了過去使用C++語言開發的成像引擎，利用GPU-OCT processing演算法對來自兩種偏振態通道的OCT訊號提供即時的處理；第二項則是針對catheter-based OCT常見的由於光點掃描樣品表面之穩定性與重複性所產生的圓周畸變(Circumferential distortion)，又被稱為非均勻旋轉畸變(Non-uniform Rotational Distortion, NURD)進行修正。在本篇論文中，共開發了兩套不同的演算法來進行NURD修正，基於NURD的特性會造成在連續時間上的影像之間有左右位移的現象，為了修正NURD，我們可以藉由找出導管遠端成像頭上金屬殼反射所引起的兩個基準標記來找到OCT影像中存在的NURD特徵。第一套演算法僅基於基準標記位置進行影像的位移以及調整，這是相對簡易的修正方式，所以有較好的速度表現。第二套演算法則是藉由分析基準標記位置隨時間的變化，可以計算出用來進行插值所需的資訊來對初始的OCT影像重採樣(Resample)以完成NURD的修正，第二套演算法的實現是基於對先前發表的版本進行改寫。兩套NURD修正演算法都使用C++語言以及NVIDIA CUDA toolkits所提供的函式庫來進行撰寫到現有的成像引擎內。兩套演算法皆會藉由NVIDIA Visual Profiler進行分析以優化資料處理及排程順序，並且在實驗室現有的catheter-based OCT系統上進行演算法搭配OCT血管攝影術(OCTA)的測試。修改過後的成像引擎能夠基於NURD修正演算法進行catheter-based OCT與OCTA成像，並且在資料蒐集完畢後可以立即進行OCT與OCTA影像的3D檢驗。我們相信在本論文裡所展示的結果將進一步促使未來catheter-based OCT與OCTA成像在臨床上應用的實現。

The development of catheter-based optical coherence tomography (catheter-based OCT), which combines the advantages OCT exhibits and also the apparatus can be used inside the human body, enables the feasibility of realizing early diagnosis of the lesions at the early stage in internal organs of the human body. However, due to the influence of the beam scanning imaging mechanism implemented in catheter-based OCT, the image acquired will be distorted. Furthermore, the presence of image distortion might lead to inaccurate computation results during advanced image analysis steps, making it challenging for catheter-based OCT to be used as a diagnostic tool in practice. In this thesis work, we have utilized graphics processing units (GPU) to develop an image correction framework for catheter-based OCT. The image correction framework is divided into two parts. The first part is to correct the artifacts caused by the change of light polarization state illuminating on the sample surface caused by fiber bending for example. In addition to utilizing polarization diversity detection module to collect inference signals from both polarization states, we have revised the imaging engine developed using C++ language previously to provide real-time processing of OCT signal from both polarization channels using GPU-OCT processing algorithm. The second part is to correct the circumferential distortion due to the stability and repeatability of the beam scanning over the sample surface, also known as non-uniform rotational distortion (NURD), commonly presented in catheter-based OCT imaging. In this thesis work, we have developed two different algorithms for NURD correction, based on the characteristics of NURD, which cause the OCT image to shift left and right, during temporal evolution. In order from correct the NURD, we can find the NURD characteristic present in the OCT imaging by identifying two fiducial markers caused by the reflection from the metal shell on the distal imaging head of the catheter. The first algorithm simply applies the imaging shifting and resizing based on the fiducial marker locations. It is a relatively simple correction method, so it has better speed performance. For the second algorithm, by analyzing the variation of the fiducial marker position as a function of time, we can compute the interpolation information to resample the initial OCT images to yield NURD-corrected ones. The implementation of the second algorithm was based on revision of a previously reported version. Both NURD correction algorithms are implemented to the existing imaging engine using C++ language and the function library provided by NVIDIA CUDA toolkits. The performance of both algorithms were evaluated using NVIDIA Visual Profiler to optimize data processing and scheduling. Then, the algorithms were tested on the existing catheter-based OCT system in our laboratory, including OCT angiography (OCTA) imaging. The revised imaging engine enables catheter-based OCT or OCTA imaging based on the either NURD correction algorithms, as well as 3D examination of the OCT or OCTA imaging data immediately after the data acquisition. We believe the results demonstrated in this thesis work should further facilitate the implementation of catheter-based OCT/OCTA imaging in the clinical practice in the future.