基於影像組學的肝纖維化超音波影像診斷系統研究

Abstract

在臨床上，醫生會使用醫學影像診斷病徵，但是診斷需要透過大量的經驗累積與訓練。為了輔助醫生診斷，目前有許多研究利用電腦輔助診斷(Computer-aided diagnosis, CAD)來分析醫學影像，利用機器學習(Machine learning)或深度學習(Deep Learning)等演算法建立病徵判別模型。然而這些模型對於判斷結果的解釋性不高，醫師難以透過模型分辨的結果來回推電腦是利用影像的哪些特徵進行判斷，因此這些模型在實際臨床使用上還難以大量投入。 在過去，肝臟超音波影像的判斷通常需要透過醫師來分析，為了協助醫生診斷病情，需要建立肝臟輔助診斷系統來提升實用性。本研究主要蒐集臨床腹部超音波影像，利用機器學習的方法，並透過影像組學方法對肝臟超音波影像進行特徵抽取，包括亮度及紋理抽取共92種特徵。先利用隨機森林(Random Forest)篩選重要性較高的特徵，再使用這些特徵重新訓練模型以分析肝臟纖維化影像組學的主要特徵特性。 研究主要訓練多種不同的隨機森林分類模型來判斷正常肝臟與患有疾病肝臟各個嚴重度分級間的區分效果，發現GLSZM(Gray Level Size Zone Matrix)和GLCM(Gray Level Co-occurrence Matrix)特徵在不同級別的肝纖維化判斷中均顯示出高度的重要性。除了分析肝纖維化的主要特徵，本研究還建立了肝纖維化分類器，以病人為單位，透過階層式分類結合多個模型，將準確率從60%提高至80%以上，以協助醫師對肝纖維化進行初步判斷。透過本研究找到的主要特徵級建立的分類模型，可提供未來醫師對於影像判讀參考的量化標準，增加檢測肝臟疾病的診斷效率以及預測準確率，建立更完善的醫療診斷系統。

In clinical practice, doctors use medical imaging to diagnose diseases, but diagnosis requires extensive experience and training. To assist doctors in diagnosis, many studies currently employ Computer-aided Diagnosis (CAD) to analyze medical images and use algorithms such as Support Vector Machines (SVM) or Deep Learning to establish disease discrimination models. For example, these models have demonstrated good accuracy in the judgment of pulmonary nodules. However, these models often lack interpretability regarding their judgment results, making it difficult for doctors to understand which image features the computer used for its judgments. Therefore, these models are still challenging to implement widely in practical clinical use. In the past, the evaluation of liver ultrasound images typically required analysis by physicians. To assist in diagnosing liver conditions and enhance practicality, this study aims to establish a liver auxiliary diagnosis system. This research primarily collects clinical abdominal ultrasound images and applies machine learning techniques combined with radiomics methods to extract features from the liver ultrasound images, including a total of 92 features related to brightness and texture. Initially, a Random Forest method is employed to select the most important features, and then these features are used to retrain the model to analyze the primary characteristics of liver fibrosis in the radiomics data. The study primarily trains multiple different Random Forest classification models to distinguish between normal liver and various severity levels of diseased liver, finding that GLSZM and GLCM features consistently show high importance in determining different levels of liver fibrosis. In addition to analyzing the main features of liver fibrosis, the study also establishes a liver fibrosis classifier. By using a hierarchical classification that combines multiple models, the accuracy rate is improved from 60% to over 80%, assisting physicians in the preliminary assessment of liver fibrosis. The classification model built using the key features identified in this study can provide quantitative standards for image interpretation, enhancing the efficiency and accuracy of liver disease diagnosis and creating a more comprehensive medical diagnostic system.