Pneumothorax Segmentation in Routine Computed Tomography Based on Deep Neural Networks

Pneumothorax is an urgent pulmonary pathology which requires diagnosis and treatment immediately or can be life-threatening. Chest computed tomography (CT) is preferred comparing with chest X-rays with respect to precision, consistency, and sensitivity to similar pulmonary pathologies. Automatically segmenting pneumothorax on routine chest CT is expected to take the place of manual segmentation since the latter is very labor-intensive and time-consuming, but the studies of the former are few now. This paper aims to apply deep neural networks to pneumothorax segmentation and illustrate some advantages and problems of this kind of deep learning method on this application. We compare the segmentation performance of ten well-known deep neural networks including nine U-Net-like models and one PSPNet model on a large and well-labelled routine chest CT dataset. Our dataset consists of 12535 CT slices from 60 pneumothorax patients. Our experiments show the potential capacity of deep neural network on pneumothorax segmentation, and we also point out some to-be-solved problems. The code is available at https://github.com/FreedomXL/Pneumothorax-Segmentation-Deep-Learning