基于改进U-Net深度网络在定量评估腕管综合征正中神经卡压中的应用

Application of improved U-Net deep network in quantitative evaluation of median nerve entrapment in carpal tunnel syndrome

目的:运用改进U-Net深度网络学习定量评价腕管综合征正中神经的超声图像,确定基于改进U-Net深度网络学习的卷积神经网络模型,探讨其在定量评估腕管综合征正中神经卡压中的应用价值。方法:搜集213例经肌电图确诊的腕管综合征正中神经卡压患者及104例健康志愿者,213例正中神经卡压患者中60例为双侧卡压。对317例受检者行超声检查,在腕管处保存超声图像,共得到正中神经图像377组。由擅长肌骨超声的医师对377组图像进行勾勒。应用基于改进U-net深度网络学习的卷积神经网络模型,分割腕管综合征卡压的正中神经超声图像,定量分析提取横切以及纵切的正中神经超声图像的影像组学量化特征。结果:改进的U-Net深度网络可以很好地识别切割正中神经;改进的U-Net深度网络可以定量表示CTS中卡压的正中神经回声减低,区域明暗参数A、明暗参数I、对比明暗参数RI以及纹理参数Homo、纹理不均匀参数Cont差异均有统计学意义(P=0.000)。结论:改进的U-Net模型在超声正中神经图像自动分割方面表现良好,可以定量分析腕管综合征正中神经卡压时灰度以及神经纹理均匀性。

Objective:To quantitatively evaluate the ultrasonic images of median nerve in carpal tunnel syndrome(CTS)using improved U-Net deep network learning,to determine the convolutional neural network model based on the improved U-Net deep network learning and to explore its application value in the quantitative evaluation of median nerve entrapment in carpal tunnel syndrome.Methods:A total of 213 patients with median nerve entrapment in carpal tunnel syndrome confirmed by electromyography and 104 healthy volunteers were collected.60 of 213 cases of median nerve entrapment were bilateral.Ultrasonography was performed on 317 subjects,and totally 377 groups of median nerve images were obtained by preserving the images in the carpal tunnel.The images of 377 groups were outlined by physicians who specialize in musculoskeletal ultrasound.The convolution neural network model based on the improved U-Net deep network learning was applied to segment the compressed median nerve ultrasonic images of carpal tunnel syndrome,and quantitative analysis was performed to extract the radiomics quantitative characteristics from the transverse and longitudinal median nerve ultrasonic images.Results:The improved U-Net deep network could identify and segment median nerve well.The improved U-Net deep network could quantitatively represent the echo reduction of the median nerve in CTS,and the differences of regional light and dark parameter A,light and dark parameter I,contrast light and dark parameter RI,texture parameter Homo,and texture heterogeneous parameter cont were statistically significant(P=0.000).Conclusion:The improved U-Net model performs well in the automatic segmentation of median nerve ultrasonic images,and can quantitatively analyze the gray scale and the uniformity of nerve texture in the case of median nerve entrapment in carpal tunnel syndrome.