NLA-3DUNet: A Non-Local Attention Mechanism-Based Network for Lung Parenchyma Segmentation

To address the challenges of insufficient integration of long-range contextual information, ambiguous boundaries, and mis-segmentation in lung parenchyma segmentation tasks, this study proposes a 3D segmentation network based on a non-local attention mechanism, termed NLA-3DUNet. Building upon the 3D-Unet+ architecture, the network embeds a channel attention module (CAM) and a position attention module (PAM) to enhance global feature perception through dynamically allocating channel weights and aggregating spatial contextual information. Additionally, the dense conditional random field (Dense CRF) post-processing step is introduced to refine segmentation boundaries. Experiments were conducted on the Parse dataset, and model performance was evaluated via five-fold cross-validation. Results demonstrate that compared to the baseline model 3D-Unet+, NLA-3DUNet achieves a dice similarity coefficient (DSC) of 0.981, reduces the volumetric overlap error (VOE) to 4.130 %, and attains a true positive rate (TPR) of 0.984, significantly outperforming comparison models such as SegNet and UTNet. Ablation experiments further validated the effectiveness of the dual attention mechanisms and deep supervision strategy. While improving segmentation accuracy, NLA-3DUNet incurs only a moderate increase in computational complexity, highlighting its high clinical application value.