Lung parenchyma segmentation based on semantic data augmentation and boundary attention consistency

The paper proposes an automatic, accurate, deep learning segmentation approach applied to the lung parenchyma. In this way, we adopt a set of strategies to produce fine and accurate segmentation. Firstly, we preprocess the image, and fill the regions outside thoracic cavity to improve the contrast of the lung parenchyma. Then, we use semantic data augmentation to get more effective dataset. Specially, we change the regions of original image inside and outside lung parenchyma except for boundary to generate new training images and preserve the invariance of lung boundary. Furthermore, we propose a novel boundary attention consistency to further enhance the attention to lung boundary and promote network to output fine boundary. We choose the UNet network as the baseline model of lung segmentation and apply above strategies to train it. The effectiveness of every component of proposed segmentation method is confirmed by ablation experiments, and the superiority of the proposed segmentation method is fully verified by comparing with other existing segmentation methods.