使用T1權重和擴散張量磁振造影影像之腦部影像模板建構

Abstract

本研究之主要目標在使用 T1 權重和擴散張量磁振造影影像來建立腦部影像模板。近年來的研究中，許多腦部影像模板是建構在一個名為 ICBM-152 的模板空間，例如：ICBM452、ICBM DTI-81 和 IIT DT 腦影像模板。然而，除了個人化變異之外，腦部結構會受種族、性別、年齡或疾病的影響而有所差異。因此，我們發展了一套系統流程，可針對所研究的族群建構其腦影像模板。為了減少在建構模板過程中所造成的影像失真，我們使用了對稱且微分同構的對位演算法，以同時提供正、逆形變場。另外，我們也提出可結合 T1 權重和擴散張量磁振造影影像資訊的目標函式，來改善影像對位的精準度。 擴散張量影像是由對雜訊敏感的擴散權重影像所估計而成。在本研究中，我們嚴謹地考量在擴散張量磁振造影影像的所有處理細節。首先，我們使用了 Medical Image Navigation and Research Tool by INRIA (MedINRIA)的方法來估計張量，此工具可適用於低雜訊比的擴散權重影像，並可確保所有估計出的張量皆為正定矩陣。為了更進一步保存所估計出張量的良好特性，我們使用了對數－歐幾里德的架構，以避免出現張量膨脹效應與非正特徵值的問題。 在此研究中，我們針對六十四個受測者的影像來建構腦影像模板。首先，我們使用剛性對位演算法將磁振造影對位到擴散權重之基準影像，以確保此二種影像對位在同一個座標空間。接著，我們在所有的影像中，找出一個對位到其他受測者影像時，擁有最小形變量的受測者作為代表。接著，我們重複地進行影像對位及逆形變場平均，直到影像模板空間收斂到一個穩定的狀態。最後，即可在此空間建構代表性受測者影像模板和平均影像模板。 本研究使用了兩種系統評估方法，其一是利用特徵值和特徵向量的組合，來評估兩個不同張量之間的重疊程度。另一個方法，則是利用擴散張量磁振造影來評估磁振造影的對位精準度。評估的結果顯示，若在非剛性對位演算法之中，同時使用 T1 權重和擴散張量磁振造影資訊，則非剛性對位的精準度可以得到改善。並且，結果也顯示我們所建立出的影像模板與對位到此模板空間的受測者影像有很高的相關性。因此，我們所提出的腦部影像模板建構流程與相關對位演算法，可以為所研究的族群提供一個腦結構分析的座標空間。

This study aims at the development of a construction algorithm for brain templates using T1-weighted and diffusion tensor (DT) magnetic resonance imaging (MRI) data. Recently, several brain templates developed in the ICBM-152 stereotactic space, such as ICBM452, ICBM DTI-81 atlas, and IIT DT brain template. In addition to inter-subject variation, however, the brain structures vary with races, genders, ages, and diseases. Hence, a construction algorithm of the stereotactic space for a specific study group can facilitate the structure analysis of the brains. Moreover, we improved the accuracy of registration procedure to reduce the image distortion during the template construction procedure. First, a symmetric and diffeomorphic non-rigid registration algorithm was used to provide both forward and inverse deformation fields. Also, we proposed an objective function which simultaneously utilized both T1-weighted and DT data to improve the accuracy of registration. The DT image is estimated from noise-sensitive diffusion-weighted images (DWIs). All details of DT-MRI processing procedure were carefully considered in this study. First, DT images were estimated from DWIs by the MedINRIA tensor estimation tool, which can tolerate the low signal-to-noise ratio (SNR) in clinical MRI and ensure the positive definiteness of all tensors. For preserving the good property of estimated tensors, Log-Euclidean metrics was used to avoid the problems of the tensor swelling effect and non-positive eigenvalues. In this study, 64 normal subjects were recruited for MRI scanning and template construction. First, we rigidly registered the MRI image to baseline DWI image for each subject to align both modalities of images in the same stereotactic space. Second, a representative subject was chosen as the one having the smallest deformation magnitude when registering to other subject images. Third, each subject image was registered to the temporary template, which was initialized as the image of the representative subject. The average of the obtained inverse deformation fields was applied to the image of the representative subject to update the temporary template. Iteratively applying the third step until the template image converges. Finally, we constructed a representative template and an average template in this converged space. In this study, two criteria were used to evaluate the constructed template images and the registration accuracy, including the DTI differences and overlaps between each subject and the template. The evaluation results showed that the accuracy of non-rigid registration was improved by simultaneously utilizing both T1-weighted and DT data. Furthermore, the results displayed a high correlation between the proposed template and registered subject images. Consequently, the proposed brain template construction could provide a stereotactic space for a specific subject group.