川崎症患者短軸心臟超音波影像之自動特徵計算與分析

Abstract

川崎症目前每年於台灣約新增800例，並已取代以往的風溼熱及風溼性心臟病成為兒童後天性心臟病的主要成因。其主要的診斷方式除了病童外觀的臨床症狀以及病理分析外，還可以透過超音波影像的表現來判斷是否為川崎症急性期，例如組織的回聲強度與血管的直徑。然而對於組織回聲強度的判斷標準，目前並沒有一致的量化方式與標準，必須經由醫務人員的經驗判斷。因此本研究希望透過一系列的流程，自動搜尋出重要的組織區域，並進一步確立能顯著判別的超音波影像特徵。 本研究所使用的自動判別流程，為先將超音波影像進行前處理，包括二值化和侵蝕與擴張，在進一步確立大動脈組織的上下範圍。接著使用計算連串長度的搜尋得到大動脈的中心點，以接著取得大動脈內腔外的四周組織像素點，最後計算所提出的可能特徵。在得到許多影像的特徵後，使用特徵選擇的方法選出顯著的特徵，並建立邏輯斯迴歸預測該張影像為川崎症患者的機率。 最終建立之邏輯斯迴歸模型，其所使用的三個特徵係數皆顯著，並且AUC達到0.881，而獨立測試的影像敏感性與特異度也到達了0.895與0.818，該模型對於辨別是否為川崎症急性期有顯著的判別能力，且本研究所提出之自動判別流程具有一定的穩定性。

Kawasaki Disease has approximately 800 new case in Taiwan annually and has replaced Rheumatic fever as the main cause of acquired heart disease in children. Other than its main diagnostic method like observation on patient’s clinical symptoms and pathological analysis, echogenicity and vessel diameter from ultrasound images can be used to determine whether the patient is in acute phase of Kawasaki Disease. However, with regard to the echogenicity in ultrasound images, there is currently no consistent quantification method or judging criteria. Thus, our research would like to propose a series of automatic detection method to search important tissue areas in heart ultrasound image and further identify significant features to determine Kawasaki Disease patients. In this research, several image preprocessing methods are applied first including binarization and erosion along with dilation. Secondly, the center of the Aorta in image is searched by calculating run lengths of pixels and then acquire pixels of tissue around the Aorta. Features proposed can be calculate after the tissue and Aortic valve are identified. Important features would be then selected by feature selecting techniques and thus logistic regression models can be constructed to predict the probability of the patient being in acute phase of Kawasaki Disease. The final logistic regression model constructed by this research use three significant features, which achieved 0.881 of Area Under ROC curve (AUC). The model also performs well in independence test with 0.842 and 0.818 in sensitivity and specificity respectively, which means the model is effective in determining acute phase of Kawasaki Disease and the quantification method is robust as well.