级联结构网络在肺癌患者颈部淋巴结超声图像分割中的应用价值

Application value of cascade structure network in ultrasound image segmentation of cervical lymph nodes in patients with lung cancer

为了准确诊断肺癌转移,本文应用深度学习技术对肺癌患者颈部淋巴结超声图像病灶区域进行分割,提出了一种用于超声图像分割的级联注意力UNet网络,该级联结构是将注意力UNet与EfficientNet相结合的二阶段分割网络,第一阶段为粗分割,第二阶段为细分割,编码器采用EfficientNet-B5作为主干网,图像多尺度输入;提出了适用于小目标、小样本场景的新损失函数;试验结果表明,本文提出的级联结构网络在肺癌患者颈部淋巴结超声图像分割中网络性能优异,Dice系数达到0.95,较其他UNet方法具有更优的分割性能。

In order to precisely diagnose the metastasis of lung cancer,a deep learning technology to segment cervical lymph node in ultrasound images of patients with lung cancer was applied in this paper,and proposed a cascade attention UNet network for ultrasound image segmentation.The cascade structure was a two-stage segmentation network combining attention UNet and EfficientNet.The first stage was coarse segmentation and the second stage was fine segmentation.The encoder adopted EfficientNet-B5 as the backbone network.The multi-scale features of the image were taken as the input.A new loss function was proposed,which was suitable for small target and few-shot scenarios.The experimental results showed that the proposed cascade structure had excellent network performance in cervical lymph node ultrasonic image segmentation,and the Dice coefficient was 0.95,which had better segmentation performance than other UNet methods.