A shape-guided deep residual network for automated CT lung segmentation

Abstract

Automatic lung segmentation is an effective method for the precise computer-aided diagnosis of lung diseases. However, CT lung scans are always complex due to issues such as weak texture, poor contrast, and variation of appearances and positions, which will affect the lung segmentation accuracy. Recently, due to strong feature expression ability, many deep convolution neural networks (DCNNs) have been proposed for application in medical image segmentation to provide an end-to-end segmentation scheme, especially the U-shape network (U-Net) and its variants. However, accurate lung segmentation methods based on DCNNs still face a certain challenge because of the insufficient process of boundary information, restricted receptive field, etc. To address these issues, with the encoder–decoder framework, a novel shape-guided deep residual network is proposed in this paper for automatic CT lung segmentation. The proposed network is composed of two stream networks: the main stream network and the shape stream network. An effective deep attention residual network is built to act as the mainstream network for lung segmentation. Meanwhile, an attention fusion block is proposed to embed into the mainstream network for multiscale feature extraction of local feature maps. Based on the mainstream network, a shape stream network is proposed to serve as significant guidance for the mainstream network to accurately compute lung shape boundaries. Multiple public CT lung image sets are adopted to qualitatively and quantitatively analyze the segmentation performance on CT scans. Experimental results indicate that the proposed shape-guided deep residual network outperforms related advanced image segmentation methods on medical image analysis.