LDUNet: a Lung Dense U-Net to segment the lung parenchyma from CT images

Automatic segmentation of lung parenchyma from computed tomography (CT) images can assist doctors in diagnosing lung cancer. The traditional lung parenchyma segmentation algorithm has poor generalization ability, and the segmentation result of CT image including pulmonary trachea and pulmonary vessels is not ideal. Hence, a LDUNet based on deep convolutional neural network is proposed in this paper for automatically segment the lung parenchyma. In order to reduce the problems of over fitting and gradient disappearance, a novel Dense block is proposed for the Encoder part of LDUNet, and the effectiveness of the Dense block is proved by ablation experiments. Secondly, SE channel attention mechanism is introduced to skip connection to reuse the high-resolution information lost in the encoding process efficiently, so that the target image generated in the decoding process is more accurate. The paper has carried out a lot of experiments based on the public LIDC-IDRI dataset, and finally achieved 99.48% accuracy rate in lung parenchymal segmentation, which is superior to most advanced segmentation models. Experimental results show that LDUNet has good accuracy and robustness in lung parenchyma segmentation.