【摘要】 针对医学图像分割任务中存在的分割目标大小变化跨度大且结构复杂，以及神经网络对目标边缘细节学习效果差这两个问题，在U-Net网络的基础上构造了金字塔空洞卷积与像素级注意力网络(DP-Net)。设计金字塔空洞卷积模块并替换传统的卷积操作，通过多种空洞卷积的组合扩展了网络感受野并编码得到全局上下文信息；提出像素级注意力模块，在通道注意力机制的基础上进一步编码像素间的依赖关系使网络能从不同通道的特征中学习到更丰富的局部上下文信息。通过在肺部公开数据集LIDC和私人肝肿瘤数据集上进行实验评估，所提出的DP-Net在三种评估指标上都获得优于当前方法的性能，证明所提出网络改进在分割精度方面的有效性。更多还原

【Abstract】 To address the problems of large variation in size and complex structure of segmentation targets and poor learning of target edge details by neural networks in medical image segmentation tasks, we propose a pyramidal dilated convolution and pixel-level attention network(DP-Net) based on the U-Net network. The dilated convolution pyramid module was constructed and designed to replace the traditional convolution operation, which extended the network perceptual field and encoded the global contextual information through the combination of multiple dilated convolutions. A pixel-level attention module was proposed to further encode inter-pixel dependencies based on the channel attention mechanism enabling the network to learn richer local contextual information from the features of different channels. Through experimental evaluation on the open lung dataset LIDC and private liver tumor dataset, the proposed DP-Net obtains better performance than current methods on all three kind of evaluation metrics, demonstrating the effectiveness of the proposed network improvement in terms of segmentation accuracy.更多还原