基于磁共振ADC图像的深度学习和ADC值评估慢性乙型肝炎肝纤维化程度价值比较

Comparison of Deep Learning Based on Magnetic Resonance ADC Image and ADC Value in Evaluating the Degree of Liver Fibrosis in Chronic Hepatitis B

目的采用基于深度学习的卷积神经网络自动检测方法对MRI的表观弥散系数(Apparent Diffusion Coefficient,ADC)图像进行分析,探究ADC图像特征与肝纤维化程度的关系。方法选取123例慢性乙型肝炎患者作为研究对象,根据肝穿刺活检,确定了5组肝纤维化阶段(F0~F4)。分别采用5层深度卷积神经网络结构对ADC图像进行分析,MRI工作站测量ADC值,2种方法进行十折交叉验证比较。结果基于深度学习采集ADC图像的准确度、敏感度、特异性、精密度、F1值(F1-Score,F1)、马修斯相关系数(Matthews Correlation Coefficient,MCC)、福尔克斯–马洛斯指数(Fowlkes–Mallows Index,FMI)分别为88.13%±1.47%、81.45%±3.69%、91.12%±1.72%、80.49%±2.94%、80.90%±2.39%、72.36%±3.39%、80.94%±2.37%,MRI工作站测量的ADC值的准确度、敏感度、特异性、精密度、F1、MCC和FMI分别为75.07%±13.35%、90.03%±9.24%、42.67%±35.42%、78.44%±11.42%、83.30%±8.18%、16.00%±60.46%、83.77%±7.98%。5层深度卷积神经网络的准确度显著高于MRI工作站测量的ADC值的准确度(P<0.01)。结论基于深度学习卷积神经网络自动检测方法的准确度优于MRI测量的ADC值,卷积神经网络自动检测方法在慢性乙型肝炎肝纤维化分期中具有较高的诊断价值。

Objective To explore the apparent diffusion coefficient(ADC)image characteristic and the degree of liver fibrosis,by the analysis of convolutional neural network based on deep learning automatic detection method for MRI of the ADC image.Methods A total of 123 patients with chronic hepatitis B were selected as research subjects.5 groups of hepatic fibrosis stages(F0~F4)were determined according to liver biopsy.The ADC images were analyzed using 5-layer depth convolutional neural network structure,and the ADC values were measured by MRI workstation,and the two methods were compared for cross-validation.Results The accuracy,sensitivity,specificity,precision,F1 value,Matthews correlation coefficient(MCC)and Fowlkes–Mallows index(FMI)of ADC images acquired by depth learning were 88.13%±1.47%,81.45%±3.69%,91.12%±1.72%,80.49%±2.94%,80.90%±2.39%,72.36%±3.39%and 80.94%±2.37%.The accuracy,sensitivity,specificity,precision,F1,MCC and FMI of ADC values measured by MRI workstation were 75.07%±13.35%,90.03%±9.24%,42.67%±35.42%,78.44%±11.42%,83.30%±8.18%,16.00%±60.46%and 83.77%±7.98%.The accuracy of 5-layer deep convolutional neural network was significantly higher than that of ADC measured by MRI workstation(P<0.01).Conclusion The accuracy of ADC value obtained by the deep learning convolutional neural network automatic detection method is better than that obtained by MRI measurement method.The deep learning convolutional neural network automatic detection method has a high diagnostic value for chronic hepatitis B liver fibrosis.