摘要 臨床上急性缺血性腦中風的早期病徵多由神經、放射專科醫師來判讀,故本研究的目的在於發展一套以電腦紋理量化為基礎之電腦輔助診斷系統,針對急性缺血性中大腦動脈中風在電腦斷層影像上的在紋理變化進行偵測及分析,利用具統計意義之特徵參數來預測早期缺血性腦中風的發生,並提供臨床醫師一套有效的輔助診斷工具。 本研究運用的技術包含影像前處理技術、類神經網路、紋理特徵分析、系統可行性評估等。首先將頭顱骨以外之雜訊予以去除,並利用所得之腦組織進行對稱軸之偵測,以分割左右腦半球。系統依據側腦室所延伸的角度,進行前大腦動脈與中大腦動脈血液分布區域之分割;而後大腦動脈與中大腦動脈血液分布區域,則依據頭顱長寬比進行分割。藉此將中大腦動脈血液分布區域予以圈選並記錄其輪廓,同時進行紋理特徵計算與擷取,如:Entropy、Energy、Contrast、Homogeneity及Low density number等。利用腦組織之對稱性質來分析左右腦半球相同血液分布區域之紋理差異變化,擷取部份具有意義之紋理參數以進行類神經網路的訓練。研究過程中,利用Leave one out 的方式進行19組案例(14組正常,5組異常)之類神經網路訓練及診斷效果評估。另外,利用ASPECTS這套視覺量化腦中風評估方式進行13組案例(8組正常,5組異常)之臨床醫師對於急性缺血性腦中風診斷能力之差異及特性分析,以驗證系統的臨床價值及其實用性。 實驗結果發現,倒傳遞類神經網路之準確率為0.95、敏感度為0.93、有效性為1.00以及Kappa value為0.88;而自我組織特徵映射類神經網路之準確率為0.74、敏感度為0.64、有效性為1.00以及Kappa value為0.49。醫師診斷測試方面發現,病變發生鑑定測試中,神經專科醫師、急診專科醫師及住院醫師之Kappa value分別為0.47、0.47及0.26。因不同科別及資歷的醫師在急性缺血性腦中風的早期診斷上確實有其差異存在,故電腦視覺紋理量化資訊能夠提供醫師視覺上額外的診斷資訊。 整體言之,本研究利用未施打顯影劑之腦部電腦斷層影像進行紋理量化分析,並提供了一套中大腦血液供應區域的分割方法,同時預測急性缺血性中大腦動脈中風的發生。系統之完成對於提供臨床醫師在急性缺血性腦中風的鑑別診斷有其正面幫助。
Abstract Early CT signs of acute ischemic stroke (AIS) are usually diagnosed by neurologists and radiologists in clinical. The purpose of this study is to develop a computed aided diagnosis (CAD) system to detect and analyze the texture changes of acute middle cerebral artery (MCA) stroke on computed tomography (CT) images based on computed texture quantification. With comprehensive texture features of the images, we could predict the risk of acute MCA ischemic stroke to provide clinicians an early detection tool of acute ischemic stroke. This study used some techniques to develop acute stroke CAD system like image pre-processing, neural network, texture analysis, and system feasibility evaluation. First of all, system will remove the useless information beyond the skull and use the symmetry information of skull stripping image to detect the delineation of hemisphere. We delineated the boundary between MCA territory and ACA territory with lateral ventricle’s angle and used the morphology of skull to delineate the boundary between MCA territory and PCA territory. This system will contour the MCA territory with saving and then calculated the texture features of MCA territory that include Entropy, Energy, Contrast, Homogeneity and Low density number . Since the characters of side to side symmetry hemispheres, we focused on the comparison in texture changes of symmetric territories by means of computed vision. System used statistical method to extract some meaningful texture features for neural network training and used the “leave one out method” for risk prediction of acute ischemic stroke. A total number of 19 datasets include 14 abnormal and 5 normal cases were trained and tested by the neural networks. Besides, clinician’s stroke diagnose trial that following ASPECTS standard to evaluate clinician ‘s diagnostic ability. The diagnostic trail of clinicians could be attest system’s feasibility and importance. Results show that the accuracy, sensitivity, specificity, and kappa value for system with Back propagation neural (BPN) classifier were 0.95, 0.93, 1.00, and 0.88, respectively. The accuracy, sensitivity, specificity, and kappa value for system with Self-organizing map (SOM) classifier were 0.74, 0.64, 1.00, and 0.49, respectively. In the result of clinician’s trial show the kappa value of neurologist, emergency physicians and residents were 0.47, 0.47, and 0.26, respectively. The difference of diagnostic experience and department will affect the result of clinicians’ trail. So quantification of computed visional textures could be provide clinician more visional diagnostic information. In conclusion, this study provide a MCA territory delineation method. The system utilize the Baseline CT to quantitate the early stage’s texture changes of acute MCA ischemic stroke and predict the occurrence of AIS . By the way, The CAD system will be able to assist clinicians and residents to diagnose the early stage of AIS.