View Identification Assisted Fully Convolutional Network for Lung Field Segmentation of Frontal and Lateral Chest Radiographs

Locating lung field is a critical and fundamental processing stage in the automated analysis of chest radiographs (CXRs) for pulmonary disorders. During routine examination CXRs, using both frontal lateral CXRs can benefit clinical diagnosis cardiothoracic diseases. However, accurate segmentation fields on still challenging due to blurry boundary poor generalization ability models. Existing deep learning-based methods focused these different type (e.g., pediatric CXRs) new diseases COVID-19) has not been tested. In this paper, view identification assisted fully convolutional network (VI-FCN) proposed simultaneously. The VI-FCN consists an FCN branch enhancing segmentation. To improve VI-FCN, six public datasets our (over 2000 were collected training. Japanese Society Radiological Technology (JSRT) dataset yields mean dice similarity coefficient (DSC) 0.979 ± 0.008, Jaccard index ( $\Omega $ ) 0.959 0.016, distance (MBD) 1.023 0.487 mm . Besides, achieves DSC 0.973 0.010, 0.947 0.018, MBD 1.923 0.755 CXRs. experiments demonstrate superior performance over most existing state-of-the-art methods. Moreover, promising results untrained COVID-19 datasets.