機器學習分析肝臟超音波影像組學的主要特徵之研究

Liver Ultrasonographic Images Analysis of Radiomics Features by Machine Learning Methods

葉俊延(Chun-Yen Yeh)

指導教授 ： 李貫銘

共同指導教授 ： 楊宏智(Hong-Tsu Young)

國立臺灣大學/工學院/機械工程學研究所/碩士(2022年)

https://doi.org/10.6342/NTU202201851

摘要

臨床上，醫師使用醫學影像診斷病徵進行診斷需透過大量的訓練與經驗累積。為了輔助醫師診斷，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，利用支援向量機(Support Vector Machine)或深度學習(Deep Learning)等演算法建立病徵判別模型；例如在肺結節判斷上有良好的正確率。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種特徵進行判斷，因此這些模型在實際臨床使用上還難以大量投入。
過去，肝臟超音波影像分析，判斷上通常需要醫師的經驗，為了提升電腦輔助診斷系統的實用性，本研究主要蒐集臨床腹部超音波影像。並利用影像組學的方法將肝臟的超音波影像進行亮度與紋理的特徵抽取，總共抽取92種特徵。先利用變異數膨脹因素的方法篩選異質性較高的特徵，再用機器學習(Machine Learning)的方式分析脂肪肝與肝臟纖維影像組學的主要特徵特性。
研究首先訓練多種分類模型分別判斷正常肝與肝臟疾病各嚴重度分級之間的區分效果。比較研究結果所利用的主要特徵是否具有解釋性，並與醫師在臨床診斷上判斷的方式進行連結。除了分析脂肪肝與肝纖維化的主要特徵，將肝臟影像的病理診斷數值化。透過本研究找到的主要特徵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，提供不同病理特徵的特徵值區間協助醫師進行不同肝臟病理的診斷，以建立更完善的醫療診斷系統。

關鍵字

超音波 ； 機器學習 ； 影像組學 ； 病理切片 ； 脂肪肝 ； 纖維化

並列摘要

Clinically, doctors need a lot of training programs and experience to make diagnosis by medical images, including ultrasonographic (US) images. In order to assist doctors in diagnosis, there are many researches using computer-aided diagnosis (CAD) to analyze medical images by using support vector machine or deep learning method. For example, these algorithms have good accuracy in the diagnosis of pulmonary nodules. However, the reason of the high accuracy result with deep learning method is hard to be understood. It is difficult for doctors to explain the results of the model resolution. Therefore, it is difficult for these models to be applied in clinical diagnosis.
In order to improve the practicability of computer-aided diagnosis system for US images, this study mainly collects abdominal US images. Using the method of radiomics, we extract 92 features from the US images of liver, including intensity and texture features. First, we use the method of Variance inflation factor (VIF) to flit the features with high independent characteristics, and then use machine learning methods to analyze the correlation among normal liver, fatty liver and liver cirrhosis features.
We use several classification models trained for liver disease, which has ability to distinguish normal liver from liver diseases. Whether the main features found in the study can be connected to the way that doctors judge in clinical diagnosis is important. The main features help quantize characteristics found in US images of liver diseases, which can provide doctors with quantitative standards for diagnosis reference in the future. This study increases the diagnostic efficiency and prediction accuracy of liver disease detection, which can help doctors to diagnose different liver diseases, so as to establish a more perfect medical computer-aided diagnosis system.

並列關鍵字

Ultrasound ； Machine Learning ； Radiomics ； Biopsy ； Fatty liver ； Fibrosis

參考文獻

[1] Farrell, Geoffrey C., “Non-alcoholic steatohepatitis: what is it, and why is it important in the Asia-Pacific region?,” Journal of gastroenterology and hepatology, vol. 18, 2, 2003.

[2] Lazo, Mariana, Jeanne M Clark., “The epidemiology of nonalcoholic fatty liver disease: a global perspective.” Seminars in liver disease, vol. 28,4, 2008.

[3] Hsu, Ching-Sheng, Jia-Horng Kao. “Non-alcoholic fatty liver disease: an emerging liver disease in Taiwan.” Journal of the Formosan Medical Association, vol. 111,10, 2012.

[4] 社團法人台灣肝病醫療策進會, “脂肪肝簡介,” .[Online]. http://www.actld.org.tw/fatty/

[5] Karen Wallace, Alastair D. Burt, Matthew C. Wright., “Liver fibrosis,” Biochem J, 2008.