Automatic airway tree segmentation from pediatric CT images with foreign bodies based on convolutional neural network

Foreign bodies in airway are most common emergency diseases in pediatrics. The peak incidence is for children of 2-3 years old, being one of the main causes of accidental death in children. In most cases, the symptoms of airway foreign body inhalation are non-specific, making early diagnosis difficult and critical. Multi-slice spiral CT examination plays an important role in the initial diagnosis and follow-up evaluation of foreign bodies in the respiratory tract. However, as the foreign body could be very small in size and varies in shape and the physicians might be stressed especially at night, it is extremely prone to missed diagnosis and misdiagnosis, affecting timely treatment and endangering the lives of children. This paper proposed to segment children's airway trees based on a deep fully convolutional neural network to aid the diagnosis. In the proposed method, the atrous convolution with different rates was employed to capture multiscale feature information and receptive fields, and the upsampling model based on global pooling and attention mechanism was explored to integrate the global content information of high-level features into low-level features. The proposed method was validated on 2 datasets. In 20 test cases of the children's thoracic data, the Dice similarity coefficient, volumetric overlap error, relative volume difference and average symmetric surface distance were 90.87%, 15.41%, 18.01% and 0. 57mm, respectively. In 8 test cases from the EXACT09 challenge, the proposed method performed better than existing methods to yield a tree length of 135. 33cm. The proposed airway tree segmentation method could be a potential tool to aid and enhance diagnosis of airway foreign bodies of children.