| 摘要: | 本研究基於機器學習運用臨床資料與舌下靜脈影像，發展電腦輔助診斷(Computer Aided Diagnosis, CAD)協助醫師判讀肝病種類。慢性肝病包含慢性肝炎、肝硬化以及肝癌，而亞洲為肝病的好發區域。實驗考慮三個二元分類問題，分別區分肝癌、肝硬化、脂肪肝與健康者。病患的臨床數據包含基本資料、抽血檢驗值、是否有抽菸、飲酒、喝咖啡的習慣；中醫相關研究指出舌下靜脈曲張與人體疾病存在關聯性，肝硬化患者的舌下靜脈曲張程度較健康者更嚴重，因此本研究除了利用臨床數據，也加入患者的舌部影像作為訓練特徵。我們使用主成分分析(Principal Component Analysis, PCA)與卷積神經網路(Convolutional Neural Network, CNN)對舌下影像提取特徵，並利用特徵選取(Feature Selection)對臨床數據篩選重要特徵，最後結合這些重要特徵作為訓練機器學習模型之用。我們利用了Random Forest、Support Vector Machines、K-Nearest Neighbors、Ridge、Logistic Regression、Multilayer Perceptron等六種機器學習方法進行實驗。實驗結果顯示，僅使用臨床資料時，機器學習對區分健康與脂肪肝者(脂肪肝：80.7%)相對於其他兩個二元分類問題(肝癌：74%,肝硬化：64.5%)有較高的準確率；另外，僅使用舌下脈絡影像當作特徵時，機器學習模型對於肝癌與肝硬化在統計有顯著差異(肝癌：52.4%、肝硬化：55.1%)，且準確率明顯低於僅使用臨床資料。最後，利用影像結合臨床資料，經特徵提取、特徵選取與網格搜尋分別在肝癌、肝硬化、脂肪肝得到準確率77.8%、70.52%、82.6%，相較於僅用臨床資料為特徵時，模型準確率提升2~6%。其中，檢測脂肪肝時在Random Forest模型下，運用CNN與反向特徵消除法(Backward Feature Elimination)，最佳準確率達85.6%；檢測肝硬化時在Multilayer Perceptron模型下，運用CNN與Ridge特徵選取法，最佳準確率達73.4%；檢測肝癌時在Multilayer Perceptron模型下，運用CNN與反向特徵消除法，最佳準確率達81.8%。現代醫學仰賴血液檢查與腹部超音波偵測肝臟疾病，我們提出的機器學習方法可以作為醫學中診斷是否有肝臟疾病的第二意見，克服僅用血液檢驗的診斷侷限。;Chronic liver disease includes chronic hepatitis, liver cirrhosis, and hepatocellular carcinoma. Also, liver disease is a severe disease, especially in Asia. In this study, we consider three binary classifications which are hepatocellular carcinoma(HCC), liver cirrhosis(LC), and fatty liver disease(FLD). All of these patients are compared with people without these diseases. There are two types of data in our data set. One is clinical or laboratory numerical data, and the other is sublingual vein imaging. We use feature selection techniques to select salient features from numerical data. Furthermore, we utilize principal component analysis (PCA) and convolutional neural network (CNN) for feature extraction from imaging. Following this, we combine these two sources of information and employ six machine learning algorithms, including Random Forest(RF), Support Vector Machines(SVM), K-Nearest Neighbors(KNN), Ridge, Logistic Regression(LR), and Multilayer Perceptron(MLP) for training.The results show that there is higher accuracy for classifying FLD(80.7%) than HCC(74.0%) and LC(64.5%) by utilizing numerical data. In addition, imaging has accuracy for classifying HCC and LC, with results as statistically significant(HCC: 52.4%, LC: 55.1%). In the case of HCC and LC, the accuracy is lower by utilizing imaging than by utilizing numerical data. Finally, we combine two data types and employ feature extraction, feature selection, and grid search in our model. The mean accuracy for HCC, LC, and FLD is 77.8%, 70.52%, and 82.6%, respectively. We improve the mean of the accuracy by 2~6% by combining features. Moreover, the best accuracy is 81.8% for HCC through MLP, CNN, and backward elimination; the best accuracy is 73.4% for LC through MLP, CNN, and RidgeCV; the best accuracy is 85.6% for FLD through RF, CNN, and backward elimination. In conclusion, in addition to blood tests and ultrasound examinations to detect liver disease, the proposed method can be potentially used as a second opinion to overcome the limitations of classical diagnostic approaches. |
