Parallel honeycomb lung segmentation network with U-shaped dual branch encoder

In order to solve the problem that existing segmentation models can hardly take into account the global context information and local detail information effectively, a two-branch parallel network incorporating the superpixel algorithm is proposed. The overall network has a U-shaped structure, and in the encoding stage, a convolutional neural network structure is used as the first branch in the encoding stage for efficient local feature extraction; the other branch uses the Transformer branch to introduce the superpixel-level self-attention mechanism to improve the ability of lesion detail information extraction. Finally, the cross-branch feature fusion module is designed to integrate the global and local semantic information acquired at each stage of the two branches to realize cross-scale feature enhancement and improve the segmentation accuracy of the network. Experiments using a large honeycomb lung dataset and a publicly available dataset demonstrate that the mIoU and mDice of this method are both improved by more than 2.8% compared to the U-Net base model.