基于注意力机制的小儿肾图分割网络

Segmentation Network of Paediatric Kidney Map Based on Attention Mechanism

摘要: 本文提出了一种全新的网络架构，解决了U-Net在小儿肾图分割中对全局信息的部分损失问题。由于U-Net的编码器通过多次下采样降低了输入图像的空间分辨率，导致模型无法完全利用图像的全局上下文信息，且下采样过程丢失了大量重要信息。因此，将Transformer架构引入U-Net的深层学习目标位置信息及局部与全局的关系。此外，设计了多尺度通道和空间注意力模块，该模块融合了多尺度特征以弥补网络池化过程中丢失的信息，能有效地捕捉了特征图的全局上下文信息，抑制背景噪声。最后，利用新华医院收集的小儿肾图数据，进行了对比实验与消融实验，验证了所提方法的有效性。

Abstract: In this paper, a novel network architecture is proposed to solve the problem of partial loss of global information by U-Net in paediatric renogram segmentation. Since the encoder of U-Net reduces the spatial resolution of the input image by multiple downsampling, it results in the model not being able to fully utilise the global contextual information of the image, and the downsampling process loses a large amount of important information. Therefore, the Transformer architecture is introduced into U-Net’s deep learning of target location information and the relationship between local and global. In addition, the Multiscale Channel and Spatial Attention (MSCA) module is designed, which incorporates multiscale features to compensate for the information lost in the network pooling process, and can effectively capture the global context information of the feature map and suppress the background noise. Finally, using the paediatric nephrogram data collected from Xinhua Hospital, comparative experiments and ablation experiments were conducted to verify the effectiveness of the proposed method.