超声影像下基于ResNet34-Unet的静脉分割方法研究

静脉超声图像存在噪点多、阈值分割效果不佳的问题，对此本文提出一种基于ResNet34主干网络的ResNet34-UNet分割网络模型，利用ResNet34网络残差学习的结构特点，在保证网络能够提取充足图像特征的前提下，有效避免梯度消失和网络退化问题，且34层的网络深度维持了较小的网络规模；利用U-Net结构特有的长连接（Skip Connection）模块，将静脉超声图像的深层特征与浅层特征有效融合，使静脉的识别精度得以较大幅度的提升，实现了静脉边缘的平滑分割。将300张静脉超声图像作为训练集，200张作为测试集，通过随机旋转、翻转、投影等操作进行数据集的增强，经过十轮迭代训练后得到模型的准确度（ACC）达96.3%，较全卷积神经网络（FCN）高5.9%，较DeepLab v3+高5.2%。结果表明基于ResNet34-UNet的静脉分割方法能够准确地分割静脉超声图像，为后续超声影像下静脉的自动识别与跟踪提供了技术参考。

Research on Vein Segmentation Method Based on ResNet34-Unet under Ultrasound Image

ABSTRACT Vein ultrasound images have the problem of too much noise and poor threshold segmentation. In this regard, this paper proposes a ResNet34-UNet segmentation network model based on the ResNet34 backbone network, using the structural characteristics of the ResNet34 network to ensure that the network can extract sufficient images. And the problem of gradient disappearance and network degradation is effectively avoided, while the network depth of 34 layers maintains a small network scale; using the unique long connection (Skip Connection) module of the U-Net structure, the deep features and shallow features of vein ultrasound images are effectively fused, so that the recognition accuracy of veins can be greatly improved, and the smooth segmentation of vein edges is realized. Use 300 venous ultrasound images as the training set and 200 as the test set, and use random rotation, flip, projection and other operations to enhance the data set. After ten rounds of iterative training, the accuracy of the model was obtained (ACC) reaches 96.3%, which is 5.9% higher than Fully Convolutional Neural Network (FCN) and 5.2% higher than DeepLab v3+. The results show that the vein segmentation method based on ResNet34-UNet can accurately segment the vein ultrasound image, which provides a technical reference for the automatic identification and tracking of the vein under the ultrasound image.