A pilot study to utilize a deep convolutional network to segment lungs with complex opacities

Lung segmentation on CT images is the first step of computer-aided system for lung diseases. Because of the complex opacities inside lung regions, the segmentation of lungs with severe diseases is not a trivial work. In this paper, we proposed a deep convolutional neural network based algorithm to segment lungs with complex opacities on CT images. This end-to-end network is composed of several convolutional layers, max-pooling layers and de-convolutional layers. We evaluated the proposed method by using 7 CT images with mild lung diseases and 42 CT images with severe diffuse lung disease including 6 kinds of complex opacities that are consolidation, ground grass opacity, honeycombing, emphysema, nodular opacity and reticular ground grass opacity. The jaccard index of the proposed method is 0.944 ± 0.066, which is higher than the one of a widely used lung segmentation method based on basic image processing techiniques [3] (0.889 ± 0.180). This pilot study gave very promising results for the utilization of deep neural network to segment pathological lungs on CT images.