DLGDFNet: a lung tumor image segmentation model based on local and global dual dynamic fusion

To address the complexity and adaptability requirements of lung tumor image segmentation, this paper proposes LGDFNet, a local–global dual dynamic fusion framework. The novelty of LGDFNet lies in the task-specific integration of input-relevant local dynamic convolution, a hybrid convolution-attention GSA module, and adaptive activation for pulmonary CT segmentation. Specifically, the local dynamic convolution branch is adapted to enhance lesion texture representation, while the GSA module combines convolutional local feature extraction with self-attention-based global dependency modeling. This design is intended to improve the joint use of local contextual information and long-range anatomical context, rather than to claim a proven advantage over window-based attention in boundary preservation. Meta-ACON is adopted as an existing adaptive activation function to improve feature nonlinearity. The proposed framework was evaluated on four public lung tumor CT datasets, including Lung-PET-CT-Dx, MSD, QIN-LungCT, and LIDC-IDRI, and the results demonstrate the effectiveness and cross-dataset robustness of this integrated design.