EA-UNet: A Macrophages Image Segmentation Model Based on U-Net with External Attention
Yi Liu, Yuanliu Nie, Ning Liu, Fengqin Yao, Jingru Zhu, Shengke Wang · 2022 7th International Conference on Image, Vision and Computing (ICIVC) · 2022
When people are exposed to air pollution for a long time, inhaling pollution particles can cause systemic diseases. Studies have shown that carbon content in Airway Macrophages (CCAM) can be used to evaluate particulate matter exposure. The automatic detection of macrophages from electron microscopy is an essential prerequisite for researchers to study the exposure of fine particles in macrophages. However, at present, the mainstream medical image segmentation algorithms are more aimed at the CT images of the lungs or brains with relatively fixed shapes, and the segmentation performance of macrophages with uncertain shapes and sizes is inaccurate. Segmentation of macrophages faces several challenges, including the high similarity of cell images leading to false detection and missed detection, and blurred boundaries in the segmentation of macrophages. Further, no public macrophage segmentation datasets are available for deep model training. To address these challenges, we propose a Macrophage Segmentation Network that combines the U-Net convolutional network as the structure of the main body network with the attention mechanism to automatically identify macrophages from induced sputum microscopy images. In our network, a parallel partial decoder is used to aggregate diverse scale features from U-Net encoder to incorporate fine-grained details with coarse-grained semantics from diverse scale features. Then, the external-attention module tokenizes image patches from the multi-scale feature map as the input sequence for extracting global contexts which focus on regions of interest to obtain more accurate segmentation results and solve the disadvantage of insufficient data. Experiments on our MASEG dataset demonstrate that the EA-UNet proposed in this paper has higher segmentation accuracy than other state-of-art segmentation methods.