Paper Title
PAU-Net: Lung Segmentation for Pulmonary Abnormality Affected Chest Radiographs

In the field of computer-aided diagnosis, lung segmentation is an important preprocessing task. Research studies have shown that by employing a lung segmentation model in an abnormality or disease detection task using chest radiographs improves the overall system performance and reliability. In this scenario, the detection accuracy becomes strongly dependent on the preciseness of the lung segmentation algorithm. In chest radiographs where due to pulmonary abnormalities lung opacities become quite dominant, the visibility of the soft-tissues decreases. As a result, the lung area and boundaries become faint. Thus, making it difficult for segmentation algorithms to estimate the structure of soft-tissues. Currently, many lung segmentation algorithms fail to address this issue leading to improper generation of binary lung masks and consequently leading to fallacious detection results. In this manuscript, the authors work towards this end and propose the - ”Pulmonary Abnormality U-Net (PAU-Net)”. PAU-Net is a deep learning based model which utilizes spatial and reverse attention layers for feature refinement and reduction. Moreover, it is trained using a hybrid loss function which enforces shape and area similarity between generated and actual lung masks. Altogether, this allows the proposed model to look for latent features corresponding to the structural pattern of the lungs in the chest radiographs. PAU-Net achieves a mean Intersection-over-Union score of 0.9242 on the testing database and outperforms other segmentation architectures - U-Net, U-Net++, Pyramid Scene Parsing Network, and Fully Convolutional Network. Keywords - Lung Segmentation, Spatial Attention, Reverse Attention, Deep Learning, Lung Opacities.