Abstract: The ground glass lesions in CT lung images are one of the important indicators for the diagnosis of COVID-19(Coronavirus disease 2019), and segmenting the lesion is critical for diagnosis, treatment, and prognosis. However, as a worldwide epidemic, heterogeneous lesions in the lungs vary greatly with the population, disease and collection equipment. It is difficult to obtain a quantitative analysis model with generalization capabilities. This paper proposes a deep convolutional network based on U-net, which has solved the problem of unbalanced foreground and background by extracting region of interesting; in addition, we integrate different resolutions networks in the framework, and achieve robust segmentation through different receptive fields. To solve the small-sample problem, data augmentation techniques are used in the training and testing stages to avoid false positives in segmentation predictions. The proposed method was evaluated on the training set and test set containing 199 and 46 multi-site CT images of patients with COVID-19 lesions. Quantitative and quanlitative evaluation suggest that the presented network could achieve superior segmentation results.