字典式壓縮感測之快速擴散頻譜磁振造影技術

Abstract

擴散磁振造影為目前臨床上唯一能造影神經纖維方向之非侵入式神經影像技術，其中，擴散頻譜磁振造影因具有解析三維空間中水分子擴散機率之能力，已被驗證能提供良好之神經纖維方向鑑別率。然而，受限於擴散頻譜磁振造影之造影原理，需沿著q空間與k空間共計六個維度進行大量之資料擷取，始能重建全腦中每個像素之水分子擴散機率分布，導致擴散頻譜磁振造影之掃描時間過長，大幅降低其臨床應用之可行性。因此，如何能夠縮短擴散頻譜磁振造影之掃描時間且使其能被廣泛應用於於臨床研究與診斷中已逐漸成為一重要之課題。

壓縮感測技術於近十年來快速發展，目前已廣泛被應用於巨維訊號處理上，例如像是資料壓縮、即時影像錄放、與無線通訊網路上。壓縮感測之原理建立於訊號本身所具備之稀疏度，由訊號所分解之基底中擷取高係數權重之係數，即可重建還原完整之訊號資料。由於壓縮感測技術能大幅縮短資料擷取之時間，並重建出近似之資料品質，已逐漸被應用於各種生醫影像技術中。

由於擴散磁振造影之q空間亦具有稀疏之特性，於應用壓縮感測於擴散頻譜磁振造影上，目前已有相關研究提出使用小波轉換原理進行資料重建，但其加速倍率受限於資料原始稀疏度。此外，先前研究亦有使用字典式基底之壓縮感測技術重建減少資料取樣之擴散頻譜磁振造影，亦能有效重建三維水分子擴散機率，並造影神經纖維之方向。雖然結合壓縮感測與擴散頻譜磁振造影已初步驗證其可行性，目前仍欠缺一有系統之分析架構以評估其效能與準確度，此外，取像參數、訊噪比及分析參數亦需進一步之探討。

此論文之主要之目的為發展快速之擴散頻譜磁振造影技術，並藉由建立字典式基底之壓縮感測技術，以重建完整之q空間訊號，並造影三維空間中水分子之擴散機率分布與方向機率分布。此外，我們提出不同之離散型字典訓練方法以及q空間資料取樣方式，以探討於不同加速倍率下之重建效能，最後並進一步探討擴散頻譜磁振造影之掃描參數對不同壓縮感測訓練模式之影響。本研究建立一套完整之量化誤差分析架構，除可提供未來於臨床應用上最佳化參數之參考外，亦希望能透過結合字典式基底之壓縮感測技術，進一步縮短擴散頻譜磁振造影於臨床系統上之掃描時間，增進其於大腦神經科學研究與臨床腦部疾病診斷之可行性。

Diffusion Spectrum Imaging (DSI) is one of the diffusion MRI techniques and has the highest accuracy of resolving complex fiber orientations in human brain. However, due to the large data sampling and resulting long scan time, its clinical feasibility has not been verified yet on clinical MRI applications. To reduce the data sampling and accelerate the scan time, a signal processing approach is highly needed without any additional cost of hardware improvement.

Compressive Sensing (CS) technique can moderate huge data information well based on the theory that extracts all the high coefficients from signal bases. This technique has been widely employed in a variety of research fields, such as data mining, wireless network communication, video and image processing. Although implementation of CS technique on DSI has been proposed in previous studies, a systematic and quantitative analysis framework is still lacking.

Therefore, this thesis aimed to establish a dictionary-based CS-DSI technique and quantitative evaluation framework. We developed a multiple-slice dictionary learning method and focused on investigating the improvement on white matter structures. We also discussed the influences of DSI sequence parameters on its performance, such as maximum b-value and signal-to-noise ratio. The framework of multiple-slice learning is verified to has higher accuracy of reconstructing probability distribution function and orientation distribution function. We expect this thesis could provide more useful information for facilitating the development of CS-DSI technology as well as utilizing this technique on neuroscience researches and clinical applications.