Lung Segmentation via Deep Learning Network and Fully-Connected Conditional Random Fields

Computer-Aided Diagnosis (CAD) benefits to early diagnosis and accurate treatment of lung diseases. As a preprocessing of CAD-based chest radiograph analysis, reliable lung segmentation is a prerequisite step which affects the precision of lesion recognition and classification. The techniques of deep learning have been widely applied for learning task-adaptive features in image segmentation. However, most existing lung fields segmentation methods based on deep learning are unable to ensure appearance and spatial consistency of the lung fields due to the varied boundaries and poor contrasts. In this study, we propose a novel method for lung fields segmentation by integrating U-Net network and a fully connected conditional random field (CRF). In the first step, we train the U-Net network designed in this paper to provide a preliminary probability to each pixel in images. Secondly, a fully connected CRF algorithm is used in this paper to optimize the coarse segmentation according to the intensity and position information of each pixel in images. Comparison with some previous methods on JSRT dataset, the proposed method in this paper shows higher Dice-Coefficient and Jaccard index.