A Lung Dense Deep Convolution Neural Network for Robust Lung Parenchyma Segmentation

Lung parenchyma segmentation is the prerequisite for an automatic diagnosis system to analyze lung CT (computed tomography) images. However, traditional lung segmentation algorithms have poor adaptability and are not effectively robust regarding lung databases with blood vessels and small voids which can interfere the segmentation. The main work of this paper is as follows: Firstly, a lung dense deep convolutional neural network (LDDNet) is proposed, which adopts some popular optimizer methods, such as dense block, batch normalization (BN) and dropout. The performance of LDDNet is tested on the public lung database LIDC-IDRI which contains many cases of interference for segmentation. Secondly, the labeled with blood vessels and small voids are not contained by the public ground-truth masks of the LIDC-IDRI database, therefore these regions are labeled by us with LabelMe software. Thirdly, for the aim of exploring the effect of image preprocessing on segmenting lung CT images with deep neural network, contrast enhancing, median filtering and Laplacian filtering are used to preprocess the image as comparative experiments. Finally, dataset is classified into four classes by the geometrical shapes to test the performance of LDDNet. The accuracy of the segmentation experiment reaches over 99% and the four classes can all reach over 95%. Additionally, blood vessels and small voids are segmented out from the lung parenchyma which is not achieved by other methods. Experimental results confirm that the proposed LDDNet can segment the lung parenchymal area more accurately and has better robustness in comparison with other neural networks and most of the traditional methods.