探討肝臟脂肪變性超音波影像於人工智能與數據挖掘方法之分類表現

Classification Performances between Artificial Intelligent and Data Mining Approaches for Ultrasound of Hepatic Steatosis

王祺元(Chi-Yuan Wang)

指導教授 ： 陳泰賓

共同指導教授 ： 杜維昌(Wei-Chang Du)

義守大學/電機資訊學院/資訊工程學系/博士(2019年)

摘要

肝臟脂肪變性已成為現今社會最常見的肝臟疾病，超音波造影不僅是臨床優先、亦是長期追蹤之診斷工具，但是其診斷結果卻受限於操作者主觀判斷。本研究的目的是利用肝臟脂肪變性超音波影像、斑塊雜訊抑制及修剪像素影像特徵，分別利用人工智能模型及數據挖掘方式，建立及探討脂肪肝超音波影像分類模型及其分類表現。
本研究採用回朔性分組實驗設計；收集共299位受試者，包括正常及輕度、中度及重度肝臟脂肪變性之超音波影像，選取感興趣區域及去除斑塊雜訊，萃取影像紋理特徵，包括灰階統計特徵、gray level co-occurrence matrix (GLCM)、gray level run-length matrix (GLRLM)、local binary pattern (LBP)、gray level edge co-occurrence matrix (GLECM)、碎形維度、肝腎比值、肝腎差值以及音波衰減率等。利用數據挖掘及人工智能方法建立肝臟脂肪變性影像分類模型；數據挖掘方法利用相關性檢測法及選用模型準確度篩選合適影像特徵，並使用簡單貝氏分類器及支持向量機分別建立數據挖掘模型。人工智能方法包括使用影像特徵建立巢狀人工智能類神經網路模型及使用肝影像建立卷積神經網路分類模型；採用測試集驗證模型之效能。
簡單貝氏分類器及支持向量機模型準確度分別為0.67及0.70，巢狀人工智能網路及卷積神經網路準確度分別為0.79及0.81。
結果顯示卷積神經網路，在分類脂肪變性超音波影像上是最可行的方法。未來最重要的課題是建立良好的訓練策略，並調校適當的參數，使得模型訓練更有效率且達到更高的準確度。

關鍵字

肝臟脂肪變性 ； 斑塊雜訊抑制 ； 巢狀人工神經網路 ； 卷積神經網路

並列摘要

Steatosis is the leading chronic hepatic disorder worldwide. Ultrasound (US) is the most utilized modality and follow-up imaging tool for visualizing liver clinically. However, the diagnostic result is often limited to operator dependency and subjective evaluation. The main aim of this study is utilizing the speckle noise suppression and trimming intensity features to build classification models for hepatic steatosis US images by artificial intelligent and data mining approaches.
In this retrospective study, hepatic US images of 299 subjects, consisting of cases of normal liver and mild, moderate, and severe steatosis, were collected. The speckle noise suppression and trimming intensity algorithm were applied to the extracted regions (ROIs) in images. Numerous features in US images, included intensity-based features, gray level co-occurrence matrix (GLCM) features, gray level run-length matrix (GLRLM) features, local binary pattern (LBP) features, gray level edge co-occurrence matrix (GLECM) features, fractal dimension, hepatorenal ratio, hepatorenal difference, and attenuation rate were utilized to build classification models using data mining and artificial intelligent approaches. In the data mining approach, significant subsets of features were selected by correlation based and wrapper subset evaluation algorithm, respectively. Naïve Bayes and Support Vector Machine (SVM) were utilized to build the classification models. Meanwhile, Nested Artificial Neural Network (Nested ANN) and Convolution Neural Network (CNN) utilized the total extracted features and hepatic images to build classification models, respectively. The training dataset was used to build models and test dataset was used to validate the performance of models.
The accuracy of classification models by Naïve Bayes and SVM classifier were 0.67 and 0.70, respectively. The accuracy of Nested ANN and CNN were 0.79 and 0.81, respectively.
The results showed that CNN was the most feasible approach to classify hepatic steatosis US images. In the future, we expect to draw up an effective training strategy and set up proper parameters which determine the efficiency and accuracy of CNN.

並列關鍵字

steatosis ； speckle noise suppression ； Nested ANN ； CNN

參考文獻

1. Fan JG. Epidemiology of alcoholic and nonalcoholic fatty liver disease in China. J Gastroenterol Hepatol. 2013; 28(1):11–7.

2. Oh MK, Winn J, Poordad F. Review article: diagnosis and treatment of non-alcoholic fatty liver disease. Aliment Pharmacol Ther. 2008; 28(5): 503–22.

3. Machado MV, Diehl AM. Pathogenesis of Nonalcoholic Steatohepatitis. Gastroenterology. 2016; 150(8): 1769–77.

4. Liu W, Baker RD, Bhatia T, Zhu L, Baker SS. Pathogenesis of nonalcoholic steatohepatitis. Cell Mol Life Sci. 2016; 73(10): 1969–87.

5. Neuschwander-Tetri BA. Non-alcoholic fatty liver disease. BMC Med. 2017; 15: 45.