ARSeg: An Attention RegSeg Architecture for CXR Lung Segmentation

Lung disease is one of the most prevalent causes of illness and death in adults and children worldwide. Researchers focused in developing computer-aided diagnosis/detection systems to provide a visual proof and to detect subtle signs of lung diseases. Various systems were proposed to separate the lung regions from organs and other tissues on Chest X-Ray images. However, there is still challenging cases such as the presence of the rib cage edges and lungs affected with abnormalities. For such, we present a new developed segmentation method called Attention RegSeg to detect the precise lung regions and improve the performance of the lung segmentation task. Attention RegSeg employs attention mechanism to suppress irrelevant features transmitted by skip connections. Experimental results showed an interesting performance with an accuracy of 95.69% and an F1-score of 95.63% using varying loss functions and three datasets (Montgomery, Shenzhen, and JSRT) as learning data. The proposed model outperformed recent state-of-the-art works and proved its ability to segment lungs overcoming challenging cases such as lungs affected with tuberculosis.