Please use this identifier to cite or link to this item:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/67964
Title: | 利用擴散磁振造影之平均表觀傳播係數分割人腦白質異常高訊號區域之演算法設計 A novel algorithm for segmenting white matter hyperintensities based on mean apparent propagator MRI |
Authors: | Chih-Hsien Tseng 曾致憲 |
Advisor: | 林發暄(Fa-Hsuan Lin) |
Keyword: | 擴散磁振造影,擴散頻譜造影,平均表觀傳播係數,白質異常高張訊號,組織分割, Diffusion MRI,Diffusion spectrum imaging,Mean apparent propagator,White matter hyperintensities,Tissue segmentation, |
Publication Year : | 2017 |
Degree: | 碩士 |
Abstract: | 研究目的:
在磁振造影T2權重影像下,人腦白質呈現較暗的灰階色彩,但因包覆白質的髓鞘易隨年齡退化,去髓鞘化的白質屬於一種白質病灶,該病灶在T2權重影像下呈現異常的高強度訊號,故又稱為白質高張訊號(White matter hyperintensities)。由於此種白質退化病徵於不同的病患間,表現的變異性大、散布範圍不一,使得尚未有一套自動化偵測技術,能定量並定性的偵測白質高張訊號區域。此外,在T1權重影像上,退化的白質會和灰質有相似的訊號表現,因此在進行組織分割(Tissue segmentation)處理時,容易造成誤判,將白質退化區域歸類於灰質組織。為了設計一能自動化標記白質異常高張訊號區域的演算法,我們利用擴散磁振造影,計算一系列平均表觀傳播係數(Mean apparent propagator index),藉由觀察白質退化區域於各系數的表現與正常組織之差異,可設定一套閥值標準,期望能在此標準下定位出白質異常區域,並經由正常白質訊號值分布作隨機抽樣,填補白質異常訊號區域,已達降低全腦組織分離錯誤率之目的。 研究方法: 在此研究中,總共使用了13名輕度認知障礙(Mild cognitive impairment)的患者,8名用於建立資料庫,藉以建立演算法,另外5名用於測試建立之演算法的偵測準確率,13名病例皆透過專業神經科學家評斷白質異常訊號嚴重程度,評斷標準遵從學界使用多年的費澤克斯分數(Fazekas score),分數範圍由0~3,分數越大表示白質退化導致得異常顯像越嚴重,此研究中使用的13名病例皆屬白質退化現象嚴重,落在分數等級3。 納入分析的影像為每人的T1權重影像以及擴散頻譜造影(Diffusion spectrum imaging),根據Özaeslan團隊的研究結果,我們利用赫米特方程式(Hermite function)計算分析擴散權重影像,並推導出11個擴散參數圖,因此所有的病例都有12種擴散參數影像(加上B0影像)。我們在8名訓練組(Training group)資料中圈選4種組織,包括腦脊液、白質、灰質以及白質退化區域,以此探測四種組織在12種擴散參數的表現,找出特定的閥值,根據此設定的閥值,可從測試組(Testing group)初步分離出四種組織,此時從全腦分離出的白值異常區域並非完整的探測,我們將此階段找出得異常區域座標,由擴散磁振影像空間轉換至T1權重影像空間,在T1影像空間以轉換後的座標當作起始點,執行區域成長(Region growing)的演算法,經過此演算法,可將原先偵測到部分的白質異常區域範圍,擴增至實際有異常訊號的白質區域,經過此步驟獲得的偵測區域,即為整套演算法所能偵測之目標區域。最終,我們透過全腦T1影像強度直方圖,找出正常白質於影像強度表現的高斯分布,進行隨機抽樣填補演算法偵測並標記為白質異常訊號表現區域,以解決T1全腦組織分離時,將白質退化區域歸類於灰質之錯誤。 由於缺少解剖數據的支持,真實白質異常區域無法有一確實的標準,用以比對演算法偵測的準確率,我們利用經驗法則,在T1權重影像上,根據影像對比差異,圈選最有可能之白質異常區域,與設計之演算法偵測到的白質異常訊號表現區域作相似性分析(Similarity index),並計算手動圈選與自動化偵測的準確偵測比例(Percentage of correct estimation)、低估比例(Percentage of under estimation)以及高估比例(Percentage of over estimation)。經過白質異常區域填補的T1影像以及原始T1影像皆會做全腦組織分離,將腦組織歸類成腦脊液、白質以及灰質三種組織,觀察其中白質組織的機率分布圖,以此評斷演算法偵測並填補白質異常區域對影像分析步驟中的組織分離,是否有提高準確率之可能。 研究結果: 我們設計的演算法對於自動化的偵測並定位白質異常區域有高度的準確率,在五名測試組病例中,手動圈選的白質異常區域與自動化偵測的白質異常區域相似性高達0.7(數值介於0~1),準確偵測的比例也高至7成,低估的比例約占2至3成,高估的比例僅1成上下。經過填補的白質高張訊號區域在影像上更趨於正常白質訊號強度,經過組織分割處理過程後,也會正確的被分配至白質類別。 結果討論: 過去的實驗接著重於利用影像對比較好的FLAIR影像,開發半自動化或自動化定位白質異常區域的演算法,然而並沒有一套能夠真正完整並完美偵測的自動化演算法,過往一般的擴散磁振造影所能提供的參數也不足以探測白質異常區域,但此研究因利用新的概念詮釋擴散磁振影像,計算出更多的擴散參數圖,提高了偵測白質異常的可能性,並提出了填補T1影像值降低組織分割錯誤的想法。總結而言,雖未達百分之百準確率,但本研究驗證了擴散磁振造影用於偵測白質異常高張訊號區域的可信度與可靠度,未來將朝向機器學習(Machine learning)的方法,提高本演算法的偵測準確率。 Introduction: White matter hyperintensities (WMHs) refer to white matter (WM) areas with increased signal intensity, appearing on T2-weighted and fluid-attenuated-inversion-recovery (FLAIR) MR images, caused by age-associated tissue decomposition in WM. Given morphological variability, scattered spatial distribution and similarity with gray matter (GM) intensity, WMHs pose great challenges in quantifying lesion volume and segmenting WM regions on T1-weighted images. Here, we proposed an automatic algorithm to identify WMHs based on microstructural indices derived from mean apparent propagator (MAP) MRI. The results of WM segmentation were compared before and after WMH localization and correction to verify the efficacy of this approach. Methods: Subjects: 13 patients with mild cognitive impairment were recruited in this research. Eight (age: 77.6 ± 4.9 years, 5 males and 3 females) were used as the training group. Five (age: 78.8 ± 9.0 years, 3 males and 2 females) were used as the testing group. Imaging: MRI scans were performed on a 3T MRI system (TIM Trio, Siemens, Erlangen) with a 32-channel phased array coil. T1-weighted imaging utilized a 3D magnetization-prepared rapid gradient echo pulse sequence: TR/TE = 2000/3 ms, flip angle = 9o, FOV = 256 × 192 × 208 mm^3, matrix size = 256 × 192 × 208, and spatial resolution = 1 × 1 × 1 mm^3. Diffusion spectrum imaging (DSI) used a twice-refocused balanced echo diffusion echo planar imaging sequence, TR/TE = 9600/130 ms, FOV = 200 × 200 mm^2, matrix size = 80 × 80, 56 slices, slice thickness = 2.5 mm. A total of 102 diffusion encoding gradients with the maximum diffusion sensitivity bmax = 4000 s/mm^2 were applied on the grid points in a half sphere of the 3D q-space with |q| ≤ 3.6 units. Analysis: According to Özaeslan’s approach, 11 MAP-MRI indices were estimated from diffusion datasets. Multiple regions of interest (ROIs) were selected in WM, GM, cerebrospinal fluid (CSF) and WMHs in training data. Values of each index in the ROIs were averaged and rescaled to characterize the index profiles of 4 different tissues. By classifying the voxels in testing group according to the index profiles generated by training group, WMHs were localized on diffusion-weighted MRI. Coordinates of voxels considered as WMHs were transformed from diffusion-weighted image space to T1-weighted image space via the transformation matrix between the two images. The voxels transformed to T1-weighted images were used as seeds for 3D region-growing to identify actual locations of WMHs on T1-weighted images. The WMH volume selected automatically by the algorithm was further compared with manual WMH mask using four similarity measures. Voxels identified as WMHs were given a proper value of signal intensity according to normal distribution of WM intensity on T1-weighted images. Tissue segmentation was performed using SPM12 and the segmentation results were compared before and after the images were processed with the correction algorithm. Results: The algorithm successfully localized WHMs on T1-weighted images. The similarity index between manual mask and automatic mask was higher than 0.7 in all the subjects. Furthermore, voxels of WMHs could be properly segmented in the tissue probability map of WM after correction. Discussion and Conclusion: This is the first study to automatically localize WMHs using diffusion parameters and correct their intensities to reduce segmentation errors. Our automatic approach combines several diffusion indices to characterize the microstructural alterations in WMHs, and has demonstrated the capability of localizing lesions and correcting segmentation errors. Future work will focus on using machine learning approach in order to improve the accuracy on WMH detection. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/67964 |
DOI: | 10.6342/NTU201701697 |
Fulltext Rights: | 有償授權 |
Appears in Collections: | 醫學工程學研究所 |
Files in This Item:
File | Size | Format | |
---|---|---|---|
ntu-106-1.pdf Restricted Access | 2.12 MB | Adobe PDF |
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.