智慧醫療：使用手部影像預測肝硬化

Abstract

肝硬化無論是在本土或是全球中皆為致死率名列前茅的疾病之一，而肝硬化傳統上的診斷方式有額外的病發症風險或是有專業人員及儀器上的要求。本篇論文嘗試透過病患身體表徵的影像結合機器學習來進行肝硬化的偵測，以達到提早發現肝硬化從而及時進行評估及治療、減少傳統方法所需的成本。 近期利用機器學習進行肝硬化預測的相關研究其資料大多須由特定儀器取得,並且僅使如支援向量機、傳統卷積神經網路等較為簡單的模型進行預測。本篇論文則希望透過較易取得之手部表面影像，並且基於肝硬化患者常有的朱砂掌病徵做出肝硬化的偵測，並且使用近期發展之影像人工智慧模型做出更精準的結果。本研究所使用的資料是由國立臺灣大學醫學院附設醫院於2020至2022年間所收集，透過篩選後所得到的1461張影像，結合專業醫療人員所標註之影像所組成之資料集。 本篇論文亦探討透過特徵擷取、特徵融合及多階段訓練等方式來進一步提升預測之準確率，並且透過實驗證實其效用。

Liver cirrhosis is one of the leading causes of mortality nationally and globally. Traditional methods to diagnose liver cirrhosis usually require specific devices and health professionals or are accompanied by complications. This thesis attempts to utilize machine learning to detect liver cirrhosis based on pictures of a patient's skin. In this way, we could detect liver cirrhosis for patients so that they could be evaluated and treated in a timely manner and the cost of detection would be reduced. Recent machine learning-based studies on cirrhosis prediction mostly require specific instruments to obtain the data. Moreover, they typically just use simple models such as support vector machines and conventional neural networks. In this thesis, we use state-of-the-art image models to achieve better prediction accuracy based on the hand palms images of the patients, which are easier to obtain. We also leverage techniques including feature transformation, feature fusion and multi-stage training to further improve our model performance. On top of that, we conduct several experiments to justify their effectiveness.