- 學位論文
結合紋理分析與影像增強技術於電腦斷層影像中輔助診斷急性缺血性腦中風
Texture Analysis and Image Enhancement for Aided Diagnosis of Acute Ischemic Stroke in CT Image
摘要
摘要 腦中風為老年人常見的疾病之一，可分為缺血性腦中風與出血性腦中風。在缺血性腦中風發病初期，主要是經由電腦斷層影像上的特徵做診斷，但由於初期特徵變化微小，造成電腦斷層攝影偵測急性缺血性腦中風的敏感度不高，無法在有效的治療時間3小時內給予治療，因而增加致殘的程度。因此，本研究的目的未整合量化分析系統與增強對比影像，實現提高電腦斷層攝影與急性缺血性腦中風之偵測率。 本研究採用急性缺血性腦中風病發3小時之內的腦部斷層影像，結合紋理量化分析以及影像增強技術進行分析。紋理量化分析：透過紋理特徵編碼中大腦血液供應區之影像並擷取其紋理特徵值，在利用腦組織具有對稱性予以分析左、右中大腦半球組織的差異性。經過t-test統計分析後，找出參數Entropy、Energy、Inertia、Intensity、Coarseness、Homogeneity、Variance以及Mean Convergence具有顯著差異。最後，採用異常18張影像以及正常4招 影像利用倒傳遞類神經網路以Leave-one-out方式進行評估。影像增強：利用5/3小波轉換，將腦部影像分解成多個頻帶。透過多值增強演算法，改善局部對比度，在經由反轉換即可得到增強後的急性缺血性腦中風影像。最後，採用10張異常與2張正常增強影像由醫師進行影像評估。 系統加入紋理特徵編碼對於急性缺血性中大腦中風的鑑別能力評估，其Accuracy由0.77提高到0.95、Sensitivity由0.78提高至1.00、Specificity則皆為0.75、Kappa value由0.40提高至0.81；很明顯看出加入特徵後對於量化系統效能更佳。醫師視覺評估結果為：Accuracy由0.63提高到0.81、Sensitivity由0.66提高至0.84、Specificity由0.5提高至0.8，病發發生位置偵測率由0.40提高至0.72；可見增強影像確能改善期鑑別能力。 本研究發展出一套具備有診斷效益的特徵參數可做為量化分析依據之系統，並提供臨床醫師視覺參考影像作為參考影像，兩者之整合可改善電腦斷層攝影對於急性缺血性腦中風偵測敏感度低的問題，並且降低誤判率。
並列摘要
Abstract Stroke, an elder general disease, includes types of ischemic stroke and hemorrhagic stroke. The diagnosis for ischemic stroke is based on computer tomography (CT) images. However, the early variation in image is too slight to detect which led hardly to treat within 3 hours and increase of lesion degree for patient. The purpose of this study is to improve the ability to detect ischemic stroke in CT images by combining analytic system of quantizing with contrast enhanced images. This study used texture quantification and image enhancement techniques to analyze the brain CT images within 3 hours after acute ischemic stroke onset. Steps of texture quantizing analysis include:1.calculated texture features of middle cerebral artery (MCA) territory after using texture feature coding method (TFC), 2.utilize the symmetric characteristics of brain to analyze the difference of hemispheres side by side. After t-test statistical analysis, some meaningful features that include Entropy, Energy, Inertia, Intensity, Coarseness, Homogeneity, Variance and Mean Converge were chosen. Finally, 18 abnormal and 4 normal cases were trained and tested thru back propagation neural network and leave one out method was separated into several sub-bands via 5/3 wavelet transform, and then use multi-scale enhancement algorithm to improve local contrast. The inverse transform method was used to create an acute ischemic stroke enhancement image in the last progress. Finally, 10 enhanced abnormal and 2 enhanced normal images were provided to physicians for evaluation. The accuracy, sensitivity, specificity and kappa value for evaluation of system to detect acute cerebral is chemic stroke with/without texture feature were 0.95/0.77, 1.00/0.78, 0.75/0.75 and 0.81/0.40, respectively. This results show that quantification analytic system has high performance after using TFCM. The accuracy, sensitivity, specificity and ratio of detected position were raised from 0.63 to 0.81, 0.66 to 0.84, 0.5 to 0.8 and 0.40 to 0.72, respectively, during the test of enhanced effect by physician. The enhanced images can raise the performance of detection for user. A system including TFCM for quantizing analysis and can provide enhancement images for clinician to diagnosis was developed. This multifunction system can improve the accuracy to detect acute ischemic stroke in CT images, and avoid the false diagnosis to patients.