肝病是台灣社會主要的文明病之一，在台灣成年男性上班族有脂肪肝的比例更高達49%。過去多以驗血方式或肝臟切片來進行脂肪肝的篩檢，但侵入式的篩檢方式，不但造成病患不適，同時也產生高昂成本，因此，具有非侵入式及便利性的肝臟超音波便成為最常用的檢測方式。然而，超音波診斷常受到醫師主觀意見影響，因此，如能提供輔助的決策判斷對醫師會有相當大的幫助。 本研究的目標在於利用卷積式神經網路（CNN）針對331位病患之肝臟超音波影像建立判斷模型並進行分類，與使用其驗血資料以機器學習方式建立之分類模型準確度進行比較。此外，本研究更進一步將機器學習與深度學習合併使用，試著找出較為適合之判斷方式。 實驗顯示以將CNN之特徵維度輸出，再以SVM進行分類，可以獲得0.821的準確率以及精確率0.862，較單純使用機器學習或是CNN來的更準確，在代替傳統侵入式篩檢方式進行輔助判別的可行性相當高。 ;Liver disease is one of the major civilized diseases in Taiwanese society. Even more, the percentage of adult male workers in Taiwan who have fatty liver is up to 49%. In the past years, blood tests and liver slices are the most often used for fatty liver screening. However, intrusive inspection methods not only cause discomfort but also high costs and potential risk to the patients. This thesis proposes a deep learning method which uses a convolutional neural network (CNN) to model and classify liver ultrasound images of 331 patients, and to compare the accuracy of classification models established by machine learning algorithms with their blood test data. Furthermore, this study tries to combine machine learning with deep learning to find a more appropriate way to judge the ultrasound images of liver. According to the experiment results, applying the SVM classification by the features extracted from CNN has better performance than using only machine learning methods. The accuracy, precision, recall and F1 score achieved 0.82, 0.862, 0.806 and 0.833 which are all better than machine learning methods with blood test data. Thus, it has a potential to diagnose fatty liver with CNN.
