LGDNet: local feature coupling global representations network for pulmonary nodules detection.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer-Aided Diagnosis

Background:

• Pulmonary nodule detection in lung CT scans is vital for early lung disease diagnosis.
• Deep learning methods show promise but struggle with small nodules due to limited long-range dependency and contextual information.

Purpose of the Study:

• To develop an advanced deep learning framework, LGDNet, for enhanced pulmonary nodule detection.
• To improve the accuracy of detecting small pulmonary nodules in lung CT scans.

Main Methods:

• Proposed a novel end-to-end framework, LGDNet, fusing local and global features.
• Designed a dual-branch module integrating CNN (local features) and Transformer (global representations).
• Incorporated an attention gate module to align and combine features from both branches.

Main Results:

• LGDNet achieved a Competition Performance Metric (CPM) score of 89.49% on the LIDC dataset.
• Significantly improved nodule detection sensitivity, especially for small nodules.
• Outperformed existing state-of-the-art nodule detection methods.

Conclusions:

• LGDNet effectively fuses local and global information for superior pulmonary nodule detection.
• The dual-branch and attention gate modules enhance the identification of small nodules.
• Demonstrates significant potential for clinical application in early lung disease diagnosis.