Multi-level supervised network for Lung segmentation

The segmentation of lung regions in CT images is an important technique in computer-assisted medical care. While convolutional neural networks have improved the ability to segment lung areas in CT images, the accuracy of current methods is limited by challenges such as low contrast and abnormal appearance. This paper describes a new approach using a multi-level supervised lung segmentation network with a U-Net backbone and a module called MSDC to address these issues. The MSDC module incorporates multi-scale dilated convolutions to extract lung edge information and address the issue of weak contrast. Additionally, the network includes spatial attention modules to improve its focus on lung areas within CT images. Using several standard CT lung image datasets, this approach demonstrated superior performance compared to recent image segmentation methods, both qualitatively and quantitatively