Dense-U-Net: A novel densely connected CNN for lung fields segmentation

Computer-Aided Diagnosis (CAD) benefits from its early diagnosis and accurate treatment of lung diseases. Accurate and reliable lung fields segmentation is a prerequisite which is convenient for further diagnosis and treatment. However, due to the superimposed regions in lung fields and varied shapes among different patients, it is difficult to ensure the consistency in appearance and spatial. Besides, while deep learning models have achieved high performance in computer vision, the large number of parameters are generated during training process, which increases memory usage. In this paper, a novel densely connected CNN (Dense-U-Net) is proposed, which is beneficial to increase information flow throughout the model network and reduce redundancies by utilizing dense connections between different layers. This network adopts two kinds of dense connections and designs a model with a smaller number of parameters. Compared with some previous methods on the JSRT dataset, the proposed method shows higher performance for lung fields segmentation with different assessment indicators.