Abstract: With the features of low Signal-to-Noise Ratio （SNR） and blurred edges， ultrasound images of the brachial plexus nerve are hard to be segmented manually. Although some results have been gained by existing segmentation models， the segmentation effect is not satisfied due to the small target area and irregular shape of the brachial plexus nerve structure. Aiming at the above problems， a multi-scale feature fusion-based brachial plexus nerve segmentation model was proposed， namely Nerve-segmentation Feature Pyramid Network （Ner-FPN）. In the feature extraction stage， an Xception-like structure was designed for multi-scale feature extraction. In the prediction segmentation stage， a bidirectional FPN structure was used for feature fusion prediction. The BP （Brachial Plexus） dataset from the Kaggle brachial plexus nerve ultrasound image segmentation competition was used as the experimental data. The experimental results show that compared with the mainstream deep learning segmentation models U-Net and SegNet （Segmentation Network），the Dice Similar Coefficient （DSC） of Ner-FPN model for brachial plexus nerve segmentation can reach 0.703， which is 10.7 percentage points and 14.5 percentage points higher than those of U-Net and SegNet， and 5.5 percentage points and 3.4 percentage points higher than those of improved models QU-Net and Efficient+U-Net in the same dataset， verifying that the proposed model can be an aid for diagnosis.

Key words: multi-scale feature, feature fusion, feature pyramid, ultrasound image, brachial plexus nerve

摘要： 臂丛神经超声影像信噪比（SNR）低、边缘模糊且人工分割难度较大。现有的分割模型虽然取得了一些成果，但碍于臂丛神经结构目标区域小、形状不规则，分割效果欠佳。针对上述问题，设计基于多尺度特征融合的臂丛神经分割模型，即针对神经部位分割的特征金字塔网络（Ner-FPN）。在特征提取阶段，设计一种仿Xception的结构进行多尺度特征提取；在预测分割阶段，采用双向FPN结构进行特征融合预测。在Kaggle臂丛神经超声影像分割竞赛的BP数据集上的实验结果表明，Ner-FPN模型对臂丛神经分割的Dice相似系数（DSC）可达0.703，与主流的深度学习分割模型U-Net、SegNet相比，分别提高了10.7个百分点和14.5个百分点，对比相同数据集中的其他改进模型QU-Net和Efficient+U-Net，DSC分别提高了5.5个百分点和3.4个百分点，可见所提模型能够起到辅助诊断的效果。

关键词: 多尺度特征, 特征融合, 特征金字塔, 超声影像, 臂丛神经