EGP-Net: a lung nodule segmentation network integrating edge guidance and pyramidal multi-scale contextual attention mechanisms.

Objectives: Accurate segmentation of pulmonary nodules in CT images is of great significance for the early screening, diagnosis, and treatment planning of lung cancer. However, manual segmentation is time-consuming and highly subjective, and existing pulmonary nodule segmentation methods still struggle to achieve accurate segmentation under challenging conditions such as blurred boundaries and interference from complex structures. This study aims to develop a segmentation method for pulmonary nodules to improve segmentation accuracy and support the clinical evaluation of lung cancer. Materials and methods: The proposed method, EGP-Net, integrates a Res2Net-50 encoder, an edge-guided network, a global pyramid perception module, a dynamic attention fusion module, and a multi-scale contextual decoder. By combining contextual information with boundary-aware features, the network can effectively represent pulmonary nodules. The model was trained and evaluated on the public LIDC dataset and a private clinical dataset, and segmentation performance was assessed using IoU, Dice, F2-score, and F0.5-score. Results: On the LIDC dataset, EGP-Net achieved an IoU of 88.32% and a Dice coefficient of 92.65%, outperforming state-of-the-art comparative segmentation methods. It also achieved excellent performance on the private clinical dataset. Ablation experiments further verified the effectiveness of each component. Conclusions: EGP-Net improves the accuracy and robustness of pulmonary nodule segmentation, facilitating precise nodule identification and quantitative analysis, and providing reliable support for lung cancer detection and evaluation.