基于U-Net改进的多尺度融合超声神经分割算法研究

An ultrasonic neural segmentation algorithm based on U-Net improved multi-scale fusion

大量传统的颈部超声神经检测算法,检测敏感性低,假阳性数量大,低层特征利用率不足。而颈部超声图像数量较少,边缘模糊且对噪声敏感。对此,提出一种改进型U-Net分支融合算法:改进损失函数,获得高质量的候选样本;使用多尺度卷积结构替换原结构中普通卷积层,增强特征提取能力;结合扩张卷积替换中、深层池化操作,提高低层特征利用率。通过对比实验验证了所提算法的算法性能。实验表明,与传统的U-Net和SegNet卷积网络对于小尺寸超声神经分割的结果相比,所提算法的分割效果较两者分别提升了近9%和17%,且对于正常尺寸和小尺寸的神经分割均有较高的分割精度。

Traditional cervical ultrasound nerve detection algorithms have low detection sensitivity,a large number of false positives,and insufficient utilization of low-level features.However,the number of ultrasound images of the neck is small,and the edges are blurred and sensitive to noise.Therefore,an improved U-Net branch fusion algorithm is proposed.It improves the loss function to obtain high-quality candidate samples,replaces the ordinary convolutional layer in the original structure with a multi-scale convolution structure to enhance feature extraction,and combines expanded convolution to replace middle and deep pooling operations so as to improve the utilization of low-level features.The performance of the proposed algorithm is verified through comparative experiments.The experimental results show that,compared with the traditional U-Net and SegNet convolution networks,the proposal improves the small-size ultrasonic neural segmentation effect by nearly 9%and 17%respectively,and the segmentation accuracy is higher for normal-size and small-size neural segmentation.