以封閉不可壓縮黏性流體模型為基礎的加速多重模式醫學影像套合

Abstract

影像套合運用在醫學研究或醫療診斷上是相當重要的技術，其目的為針對一系列相關影像，透過尋找兩影像間空間變換，將一張影像映射到另一張影像，並將個別的資訊顯示在套合後的影像。本論文使用物理模型類中的黏性流體方法，前人開發的演算法在求解納維-史托克方程式及對形變場雜訊做濾波時相當耗時，並且無法進行多重模式醫學影像套合，故本研究針對這些問題進行改良。在本論文中，我們以雅可比(Jacobi)方法疊代求解統御方程式中的隱性黏滯項，並利用GPU強大的平行處理能力，搭配NVIDIA開發的CUDA架構來平行加速前人演算法中數個步驟。另外我們利用互資訊改良物體力方程式，完成多重模式醫學影像套合。最後我們使用了多組不同類型的磁振影像來評估此方法，實驗結果顯示，本研究提出的方法可有效處理多種類型的影像套合，包含：去頭殼影像、雜訊影像、大規模變形影像及多重模式影像，本方法不僅成功降低一半處理時間，達到2.2~2.6倍的加速，套合後影像在相關係數、差方和上亦有良好表現。

Image registration is an important technique for medical research and medical diagnosis. It is a process of looking for a spatial transformation between two images and mapping one to the other one based on the transformation function. There are many image registration methods and one algorithm is based on a non-rigid fluid flow model. However, the computation of the governing equation and Gaussian smoothing of this method is quite time-consuming and it is unable to perform multimodal registration. To address these problems, we adopt the Jacobi method iteratively to solve the implicit viscosity terms and parallelize the program with GPU. Compute Unified Device Architecture (CUDA), an application programming interface for GPU by NVIDIA, is used to accelerate the algorithm. Besides, we modify the body force term via the mutual information to achieve multimodal image registration. A variety of different types of magnetic resonance images were used to evaluate this new method. Experimental results indicated that the proposed method efficiently registered many kinds of images, including skull-stripping images, noisy images, large scale deformation images and multimodal images. Comparing to the previous fluid-flow model, our method approximately reduced the processing time by half and successfully achieved multimodal image registration.