Fu T T, Yao Z, Ding H, Xu Z T, Yang M R, Yu J H, Wang W P
Department of Ultrasound, Zhongshan Hospital, Fudan University, Shanghai Institute of Medical Imaging, Fudan University, Shanghai 200032, China.
Department of Electronic Engineering, Fudan University, Shanghai 200433, China.
Zhonghua Yi Xue Za Zhi. 2019 Feb 19;99(7):491-495. doi: 10.3760/cma.j.issn.0376-2491.2019.07.003.
To establish automatic liver fibrosis classification models by using traditional machine learning and deep learning methods and preliminaryly evaluate the efficiency. Gray scale ultrasound images and corresponding elastic images of 354 patients, 247 males and 107 females, mean age (54±12) years undergoing partial hepatectomy in Zhongshan Hospital of Fudan University from November 2014 to January 2016 were enrolled in this study. By using traditional machine learning and deep learning methods, an automatic classification model of liver fibrosis stages (S0 to S4) were established through feature extraction and classification of ultrasound image data sets and the accuracy in different classification categories of each model were calculated, by using liver biopsy as the reference standard. Pathological examination showed 73 cases in pathological stage S0, 40 cases in S1, 49 cases in S2, 41 cases in S3, and 151 cases in S4. The traditional machine classification model based on support vector machine (SVM) classifier and sparse representation classifier and the deep learning classification model based on LeNet-5 neural network, their accuracy rates in the two categories (S0/S1/S2 and S3/S4) were 89.8%, 91.8% and 90.7% respectively; the accuracy rates in the three categories (S0/S1 and S2/S3 and S4) were 75.3%, 79.4% and 82.8% respectively; the accuracy in the three categories (S0 and S1/S2/S3 and S4) were 79.3%, 82.7% and 87.2% respectively. Computer-aided assessment of liver fibrosis progression in patients with chronic hepatitis B has a high accuracy, and can achieve a more detailed classification. This method is expected to be applied in the non-invasive evaluation of liver fibrosis in patients with hepatitis B in clinical work in the future.
运用传统机器学习和深度学习方法建立自动肝纤维化分类模型，并初步评估其有效性。本研究纳入了2014年11月至2016年1月在复旦大学附属中山医院接受部分肝切除术的354例患者的灰度超声图像及相应弹性图像，其中男性247例，女性107例，平均年龄（54±12）岁。通过传统机器学习和深度学习方法，对超声图像数据集进行特征提取和分类，建立肝纤维化分期（S0至S4）的自动分类模型，并以肝活检为参考标准计算各模型在不同分类类别的准确率。病理检查显示，病理分期S0为73例，S1为40例，S2为49例，S3为41例，S4为151例。基于支持向量机（SVM）分类器和稀疏表示分类器的传统机器分类模型以及基于LeNet-5神经网络的深度学习分类模型，其在两类（S0/S1/S2和S3/S4）中的准确率分别为89.8%、91.8%和90.7%；在三类（S0/S1和S2/S3和S4）中的准确率分别为75.3%、79.4%和82.8%；在三类（S0和S1/S2/S3和S4）中的准确率分别为79.3%、82.7%和87.2%。计算机辅助评估慢性乙型肝炎患者肝纤维化进展具有较高的准确性，且能实现更细致的分类。该方法有望在未来临床工作中应用于乙型肝炎患者肝纤维化的无创评估。
