基于改进DeepLabV3+的COVID-19肺部CT图像语义分割方法

Semantic Segmentation Method of COVID-19 Lung CT Images Based on Improved DeepLabV3+

摘要: 肺部CT图像是确诊新冠肺炎的必要参考，但其人工判读的效率较低且专家不足。对CT图像进行高效、准确分割是(半)自动诊断的基础，本文基于新冠肺炎患者肺部CT图像及其掩膜图构建了数据集，根据实验选用和改进DeepLabV3+模型，包括在特征拼接中引入卷积注意力模块，用深度可分离卷积替换常规卷积，获得了实验性能突出的语义分割改进模型。

Abstract: CT images of the lungs are necessary references for diagnosing COVID-19, but the efficiency of manual judgment is low and there are insufficient experts. In order to use deep learning to improve its diagnostic efficiency and accuracy, this paper constructs a data set consisting of real lung CT images of patients with COVID-19 and corresponding masks. According to the experiment, the DeepLabV3+ model is selected, we introduced convolutional attention module in the feature stitching and used depth separable convolution to replace the original conventional convolution. Finally, we obtained an improved semantic segmentation model with outstanding experimental performance.