Lung Parenchyma Segmentation Based on U-Net Fused With Shape Stream

Abstract

Accurate lung parenchyma segmentation is vital for computer-aided lung cancer diagnosis. Existing lung parenchyma segmentation networks excel at segmenting large and clear lung parenchyma regions, but struggle with small and blurry regions. This study proposes an improved network structure to improve the segmentation performance of small and blurry lung parenchyma regions while maintaining accuracy for large and clear lung parenchyma regions. The proposed network is an improved network based on U-Net. A shape stream branch and multi-scale convolutional blocks are introduced into the network. The proposed network takes computed tomography (CT) images as inputs and generates corresponding binary masks as outputs. In this study, the original CT images are derived from the Open Source Imaging Consortium (OSIC) Pulmonary Fibrosis Progression dataset. In the experimental results, the overall mean Dice Similarity Coefficient (DSC) of the proposed network reaches 0.932907, which is 0.64% higher than that of the second-best network. The segment test results show that the proposed network has significantly better segmentation performance than other networks when the lung parenchyma is relatively small. On a patient-by-patient basis, the mean DSC of the proposed network is 0.63% higher than that of the second-best network. The proposed network improves the segmentation performance of small and blurry lung parenchyma while maintaining accuracy for large and clear lung parenchyma.

Keywords

• Deep learning
• lung parenchyma segmentation
• shape stream
• multi-scale feature