MSE-UNet:基于多尺度特征和注意力机制的细胞分割模型

MSE-UNet:cell segmentation model based on multiscale features and attention mechanism

细胞图像的自动分割是目前非常困难且具有挑战性的问题.在U-Net网络的基础上,增加了多尺度跳跃连接,引入了通道注意力模块,提出一种改进的网络模型MSE-UNet(multiscale skip connection-squeeze and excitation-UNet).多尺度跳跃连接将不同层次的信息巧妙地结合在一起且避免了冗余,能更加精确地分割出目标,有效地解决了细胞边界不明确的问题.通道注意力模块学习各个通道的重要性,从而使重要的特征通道占据更大的比重,在最终呈现的输出图像中展现出分割网络重点关注的部分,有效解决背景杂乱的问题.在两个公开数据集和一个自建数据集上进行测试,实验结果显示,与最近几年提出的细胞分割模型相比,该模型具有更好的性能.

This paper proposes an improved network model named MSE-UNet(multiscale skip connection-squeeze and excitation-UNet),based on the U-Net network,a multiscale skip connection is added and a channel attention module is introduced.Multiscale skip connection skillfully combines different levels of information and avoids redundancy,accurately segments the target,and effectively solves the problem of unclear cell boundary.The channel attention module learns the importance of each channel to make the important feature channels account for a larger proportion.As a result,the key part of the network will be shown in the final output image,which effectively solves the problem of background clutter.The experiment results on two public data sets and one self-built data set show that this proposed model obtains better segmentation performance.