Ultrasound multifeature classification of fatty liver disease using random forests

• Main Authors: Hsi-Shen Chen, 陳希聖
• Other Authors: 張建成
• Format: Others
• Language: zh-TW
• Published: 2019
• Online Access: http://ndltd.ncl.edu.tw/handle/j3b4ya

碩士 === 國立臺灣大學 === 應用力學研究所 === 107 === In recent years, liver disease has become one of the main diseases of Taiwan, and its priority risk factor fatty liver disease is gradually taken more seriously. Fatty liver disease is in a reversible path in the early stage, but if it goes into the later stage of fibrosis, it may even cause cirrhosis, therefore early diagnosis and treatment are particularly important. The current gold standard for fatty liver diagnosis is liver biopsy. However, it is impractical as a diagnostic tool for it is an invasive diagnostic method. In other imaging methods, ultrasound is the best way to diagnose fatty liver because of its non-invasive, non-radioactive, reusable and low cost. However, due to the characteristics of ultrasonic imaging in time, it is necessary to have well-trained personnel in operation and the experience gap between observers will lead to different ultrasonic diagnosis results. Thus quantitative ultrasonic method was created. This study aims to use three ultrasound features representing different meanings are Shannon entropy (SE), which represents the microstructure change of the liver parenchyma in the ultrasound image; the attenuation coefficient (AE) is a quantifiable coefficient of the attenuation of the sound wave in the medium; Integrated backscatter (IB) is a function that can represent the average power. These three ultrasound features are combined with the three characteristics commonly used by doctors to determine fatty liver. Body Mass Index (BMI) and Aspartate Transaminase (AST), Alanine transaminase (ALT), to assist doctors in the diagnosis of fatty liver. This study used random forest algorithm in machine learning, combined with the six features mentioned above to train a random forest model to determine 5% fatty liver patients and 33% fatty liver patients, and finally reached a 5% binary classification model 80.3% accuracy and the 33% binary classification model achieved an accuracy of 90.1%. In the three classifications, the accuracy of directly training a multiclass model reached 68.8%, while the accuracy of the multiclass model by using successive dichotomies can be improved to 72.1%.