LSEG: Lung Segmentation for Pulmonary Disease Affected Chest Radiographs

Approximately four million people die due to lung diseases. A shortfall of 42% pulmonologists is estimated by 2030. This is an indication of overburden on the pulmonologists and delay in pulmonary disease diagnosis. Computer vision-based tools may prove useful in addressing these challenges by assisting health experts in the primary screening of lung abnormalities. However, the extraction of irrelevant features may degrade the reliability of these techniques. This raises the need for employing pre-processing techniques. Many researchers employed lung segmentation techniques for the preprocessing of chest radiographs. These techniques lack in identifying soft tissues and lung boundaries from chest radiographs of infected people. This is due to an increase in opacity due to lung infection or abnormality. In this manuscript, the authors focus on recognizing soft tissues and boundaries even in infected radiographs. They propose a lung segmentation model (LSEG) that utilizes spatial and reverse attention layers for discarding irrelevant features and extracting relevant features. It employs a loss function derived using binary cross entropy, mixed gradient, and intersection-over-union loss functions. This enables LSEG to locate the structural pattern of the lungs in the chest radiographs. The model achieves the highest accuracy of 97.12% on the testing dataset and outperforms other segmentation architectures such as U-Net, U-Net++, Pyramid Scene Parsing Network, and Fully Convolutional Network.