慢性肝病常會進行到不可逆的肝硬化階段，但是在肝纖維化的早期，其病勢是可逆的。 雖然侵入式診斷是目前評斷肝纖維化的黃金標準，除了成本較為高昂且會造成病人的 痛苦且帶有些微機率的後遺症。所以我們希望能建立一套簡單但可靠的非侵入式評分 方法，並期望能判別患者肝纖維化的程度。主要影像是使用小鼠的MR 影像，利用 MR 影像的T1 和T1ct (Primovist) 各自的成像差異加上K-means 法把肝影像分割出來。我們架構的模型主要使用Support vector machine (SVM) 去進行分類。透過觀察,我們使用simple connected domain 的機率當作其中的一個特徵，並用孔隙率當作第二個特徵。由於SVM 本身是Binary 分類器，所以我們選用One-Against-One 策略，使用多個Binary 分類器去達到我們想要的分類效果。由於資料組數較少，因此我們選用K-Fold 交錯驗證法來增強結果的可信度。單獨使用孔隙率當作特徵所得到的模型準確率大約有80%，根據ROC 曲線分析，可以得出其AUROC 分數大約為0.7 ∼ 0.8 之間，但如果加上simple connected domain 作為第二特徵，訓練出來的模型準確率可達 90%，AUROC 分數可達將近0.9，代表其判斷結果是可信的。;Chronic liver disease often progresses to the irreversible stage of cirrhosis, but in the early stage of liver fibrosis, its disease is reversible. Although the invasive diagnosis is currently the gold standard for judging liver fibrosis, in addition to the high cost and the painful sequelae of the patient with some micro-probability. Therefore, we hope to establish a simple but reliable non-invasive scoring method and hope to determine the degree of liver fibrosis in patients. The main image is the MR image of the mouse. The liver image is segmented using the respective imaging differences of T1 and T1ct (Primovist) of MRI plus the K-means method. The model of our architecture mainly uses a support vector machine (SVM) to classify. Through our observation, we use the degree of the simple connected domain as one of the features, and the porosity as the second feature. Since SVM is a binary classifier, we choose the one-against-one mode and use multiple binary classifiers to achieve the classification effect we want. Due to the small number of data sets, we choose K-Fold cross-validation to enhance the credibility of the results. The accuracy of the model obtained by using porosity alone as a feature is only about 80%. According to the ROC curve analysis, it can be concluded that its AUROC score is about 0.7∼0.8, but if the simple connected domain is added as the second feature, the accuracy of the trained model can reach 90%, and the AUROC score can reach close to 0.9, which means that the judgment result is credible.
