腦中風高居全國十大死因的前三名,台灣地區最常見的腦中風是缺血性腦中風。若能以最快的速度以導管手術打通堵住的血管,則癒後會最好。在急性缺血性腦中風的電腦斷層影像中,會呈現高亮度血管特徵(Hyper -dense Artery Sign, HAS) 。而肉眼辨識HAS特徵在0.6mm細切的頭部CT影像辨認率較佳,但其偽陽性與偽陰性的機率仍偏高,且人眼辨識速度較慢。此論文主要為開發一套急性缺血性腦中風血塊之電腦偵測診斷系統,來輔助醫生診斷時,提高其辨識HAS的準確率及效率。 此研究是透過影像處理技術來建構此系統及平台並以不使用顯影劑之細切腦部電腦斷層影像為標的。首先以圈選ROI來得到中大腦動脈區域,再對得到的中大腦動脈區域透過影像前處理、閥值分割及形態學處理取得中大腦動脈影像,再從中大腦動脈影像找出可能為HAS的特徵參數。隨後,以支持向量機訓練出適用辨識HAS之特徵參數。此研究使用了既有之8組CT細切影像來進行測試及訓練,最後請醫師以肉眼判讀及專業知識判讀HAS特徵及診斷結果與本系統之結果驗證。所開發的平台之功能包括影像讀取、區域圈選、中大腦動脈分割及參數分析。 研究結果顯示利用中值濾波及臨界值法可將ROI中之中大腦動脈區域有效分割出,以形態學處理與相連腦組織分開,在紋理特徵參數統計後結果顯示在Entropy、Energy、Contrast、Inertia等四個參數有統計上之顯著差異。根據該參數能有效找到HAS,評估後系統之效能指標Accuracy 、Sensitivity、 Specificity、及Kappa value分別為82.5%、90%、80%、及0.65。 本研究所開發之系統經初期資料評估為可實用之系統,可以將中大腦動脈上HAS特徵與正常血管區分。未來在經較大量資料之驗證後,將可提供對於急性缺血性腦中風初期血塊偵測研究有正面的幫助。
Cerebrovascular disease ranked third among the top 10 leading causes of death in Taiwan and the most common type of stroke is ischemic stroke. If the clot in blood vessels by intra-arterial thrombolysis can be removed rapidly, the prognosis will be better. A Hyper-dense Artery Sign (HAS) can be found in the CT image of acute cerebral Infarction. And the recognition rate of HAS in 0.6mm slice CT is better than common slice by the naked eye. However, by using naked eye recognition method, the rate of false positive and false negative is high and spends a lot of time. In this study, a computer-aided detection (CAD) system that can help doctor to improve the accuracy and efficiency for HAS recognition was developed. In this study, digital image processing methods are used to construct the system for 0.6mm slice CT to analysis the HAS. At first, the ROI was circled to get the MCA region and then the MCA image will be segmented by using preprocessing, single thresholding, and morphology methods. Texture and feature parameters were analyzed for MCA image to find the suspected HAS region. After evaluating the results by the independent t-test, the effective features were selected and severed as inputs in the support vector machines (SVM) to classify normal and HAS. In this study, 8 groups of thin slice brain non-contrast CT image from image database are used for the training and testing system. The system is evaluated by compared the result of specialist physicians in naked eye. The system panel with functions of image input, ROI selects, MCA segmentation, and texture parameters analysis was developed as well. The results show that MCA was separated well by median filter, thresholding, and morphology in the ROI. In the texture analysis part, the result of statistic show that Entropy, Energy, Contrast, Inertia were statistically significant in artery. According to the parameters, the performances of system in accuracy, sensitivity, specificity, and kappa value were 82.5%, 90%, 80%, and 0.65, respectively. The developed system in this study has been evaluated as a practical system based on the 8 groups images. In the future, after the verification of a large amount of images, this system will be positively helpful for detecting hyper-dense clot detection in acute cerebral infarctions research.