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

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer-Aided Diagnosis

Background:

• Accurate pulmonary nodule segmentation in CT images is crucial for lung cancer management.
• Manual segmentation is subjective and time-consuming, while existing methods face challenges with blurred boundaries and complex structures.

Purpose of the Study:

• To develop an improved segmentation method for pulmonary nodules in CT images.
• To enhance segmentation accuracy and robustness, supporting clinical lung cancer evaluation.

Main Methods:

• A novel method, EGP-Net, was developed, integrating Res2Net-50, an edge-guided network, global pyramid perception, dynamic attention fusion, and a multi-scale contextual decoder.
• The model combines contextual information with boundary-aware features for effective pulmonary nodule representation.
• Trained and evaluated on public (LIDC) and private clinical datasets, with performance metrics including IoU, Dice, F2-score, and F0.5-score.

Main Results:

• EGP-Net achieved high performance on the LIDC dataset with an IoU of 88.32% and Dice coefficient of 92.65%.
• The method outperformed state-of-the-art segmentation techniques and demonstrated excellent results on a private clinical dataset.
• Ablation studies confirmed the effectiveness of individual components within EGP-Net.

Conclusions:

• EGP-Net significantly improves the accuracy and robustness of pulmonary nodule segmentation.
• The method facilitates precise nodule identification and quantitative analysis.
• EGP-Net offers reliable support for lung cancer detection and clinical evaluation.