Towards A Relation Extractor Nested U-Architecture for Accurate Pulmonary Airway Segmentation in CT images

Pulmonary airway segmentation is essential to computer aided diagnosis and surgical planning systems for pulmonary diseases. Unfortunately, it usually gets trapped in voxel leakages at the peripheral (small) bronchi due to the partial-volume effect. To this end, we present a new deep learning relation extractor nested U-architecture that combines convolution and a self-attention mechanism in an encoder-decoder mode. Specifically, we employ convolution in the shallow layers to extract the feature of the single trachea and bronchi in a short range and introduce self-attention in the deep layers to capture the whole airway-tree structure in a long range. Both convolution and self-attention operations are embedded into an U-architecture. We evaluate our deep learning architecture on 30 computer tomography volumes, with the experimental results showing that our method can improve the average coefficient dice to 0.884 and reduce the false positive rate to 0.019.