Att GGO-Net:A Semantic Segmentation Method of Lung CT Images with Self and Cross Attention Mechanism

Abstract

Chest computed tomography imaging (CT) has become a diagnostic method for many lung diseases due to its high sensitivity and low missed rate of infection. Convolutional neural network based semantic segmentation model that can assist physicians in auxiliary diagnosis. The existing semantic segmentation models do not work well for lesion segmentation of lung CT images, small lesions are not sensitive, and the contour segmentation is coarse, for which this paper designs a network named Att GGO-Net specifically for lung CT images ground glass opacity(GGO) and Lung consolidation segmentation.The basic encoder-decoder network structure used in this study is the U-Net++ model. After feature mapping is finished at each level of U-Net++, cross attention mechanism is introduced to optimize the semantic information in the decoder layer. Finally, self attention mechanism is added to complete the overall network structure.The extensive attention mechanism will cause the network to have many parameters. To solve the issue of large network parameters, the research substitutes depthwise separable convolution for traditional convolution. The data collection of lung CT scans with ground glass opacity and lung consolidation was used for the experiments. The segmentation performance was compared between Att GGO-Net and three classical medical image segmentation networks, namely, U-Net, U-Net++, and Attention U-Net (Att U-Net). The experimental findings demonstrate that the strategy used in this research has a superior segmentation effect than the comparison method.