利用圖形處理器加速傅立葉式光學同調斷層掃描術之即時化影像處理

Abstract

在本篇論文中，我們使用了圖形處理器(GPU)加速傅立葉式光學同調斷層掃描術(FD-OCT)的影像重建。首先、我們透過AlazarTech提供的應用程式介面編撰加速OCT影像處理的程式，並與我們實驗室所開發的使用者介面結合。透過將其與現有多執行緒的程式做影像重建的速度比較，發現使用GPU在影像處理的速度上有顯著的提升。再者，我們透過鑷子在牙齒樣本上的移動模擬手術的過程，即時影像的呈現相當流暢。即便AlazarTech提供很方便的開發工具，但針對需要波長校正的光源並沒有相對應的函式，且實驗室未來需對GPU上的資料作更進一步的處理。基於這些原因，我們開發了利用統一整合架構的影像處理方法，透過平行運算的方式加速OCT影像處理，並且用NVIDIA Visual Profiler對程式進行分析及優化，使其在影像處理的表現與AlazarTech所提供的函式是可媲美的。最後我們比較幾種常見的內插方式在計算效率以及影像品質的表現，選定三次樣條插值(Cubic spline interpolation)做為波長校正的方式。在未來的研究方向，可以利用目前所發展的程式對空間頻域式光學同調斷層掃描術(SD-OCT)進行影像的處理，甚至可以利用GPU進行三維OCT影像的視覺化或是進行組織分類。

In this thesis work, we have utilized the graphics processing unit (GPU) to accelerate the reconstruction of the Fourier-domain optical coherence tomography (FD-OCT) images. First of all, we utilized the application programming interface (API) functions provided by AlazarTech Inc. to develop a GPU-based OCT processing method. Then, we integrated this processing method with our in-house developed graphic user interface (GUI). To demonstrate the feasibility of our work, we compared the GPU-based processing method with our existing multithreading processing method. The result shows that the processing speed has been substantially enhanced. Moreover, we have demonstrated the imaging of moving a tweezer on the tooth specimen to simulate the surgery, and the instant preview has displayed smoothly without delay. Even though AlazarTech Inc. provided us a convenience development toolkit, the API functions do not support the resampling process, essential for the OCT system requiring wavelength calibration procedure. Besides, in the future, we aim to apply an additional imaging process algorithm to the reconstructed OCT images with GPU. Owing to these reasons, we would like to develop a GPU processing method based on Compute Unified Device Architecture (CUDA) to manage the data on GPU more efficiently. We used the NVIDIA Visual Profiler to analyze the efficiency of the developed program and tried to optimize its performance comparable to the computation performance with AlazarTech GPU APIs. Lastly, we compared three different interpolation methods with their image quality and computing efficiency, and we eventually decided to use cubic spline interpolation as the default wavelength calibration algorithm. In future work, we can acquire and perform real-time processing on the data acquired with spectral-domain OCT leveraging the framework developed in this thesis. Moreover, we can display the visualization of the three-dimensional (3D) OCT images or compute some massive calculations work, such as tissue segmentation, with the support of the GPU.