MSU-Net: automatic segmentation of lung infection boundaries and small lesions in CT images

Lung Computed Tomography (CT) images play an important role in the diagnosis and treatment of patients with COVID- 19. However, manually identifying lesions in CT images is very time-consuming and requires specialized medical knowledge. The accuracy and integrity of the existing models are still not ideal due to the reasons such as scattered distribution, small lesions and blurred boundaries. Therefore, we propose a segmentation model based on U-Net architecture by combining Mirror-symmetry Boundary Guided (MBG) module and then adding Spatial Attention Dilation (SAD) convolution Module, called MSU-Net. The SAD module uses spatial attention mechanism and dilated convolution and multi-scale feature extraction strategy to enhance the network's ability to recognize small lesions. The MBG module further enhances the model's ability to capture more complex context information and boundary details, making the model more robust and able to effectively deal with the challenges of fuzzy boundaries. The proposed method has shown superior performance in terms of the Dice coefficient, Hausdorff distance, and Sensitivity on the publicly available COVID-19 CT dataset.