Application of UNet++-based automatic segmentation model of COVID-19 lesions in CT slices

In view of the diversity of lung lesions and the complexity of regional segmentation, a new UNet++ model is proposed, including improvements on the basis of U-Net, mainly composed of extrusion and attention modules, empty space pyramid pooling module, down It consists of sampling, upsampling, skip connections and loss functions. First, squeezing and attention modules are introduced to strengthen the attention of pixel grouping, make full use of global context information, and allow the network to better mine the differences and connections between pixels. Second, a hollow spatial pyramid pooling module is designed to capture multi-scale information of COVID-19 lesions. Downsampling obtains high-dimensional information, then four upsampling is used to restore the feature maps to the original size, and four skip connections are used to merge the feature maps. In addition, the generalized Dice loss can reduce the correlation between lesion size and Dice loss, thus solving the small region segmentation problem. Extensive experiments are performed on the UNet++ model using CT scan data from different datasets. In experiments, the UNet++ model and GDL are compared with typical segmentation models and popular loss functions, respectively, and the experimental data show that the proposed new UNet++ model is closest to the gold standard.