MSANet: Multiscale Aggregation Network Integrating Spatial and Channel Information for Lung Nodule Detection

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer-Aided Diagnosis

Background:

• Accurate pulmonary nodule detection is crucial for early lung cancer diagnosis and treatment.
• Existing methods face challenges in effectively extracting and fusing multiscale features.

Purpose of the Study:

• To propose a novel multiscale aggregation network (MSANet) for enhanced 3D pulmonary nodule detection.
• To improve the accuracy and efficiency of identifying pulmonary nodules in medical scans.

Main Methods:

• Developed MSANet integrating spatial and channel information for multiscale feature fusion.
• Employed multiscale aggregation interaction strategies to handle resolution differences.
• Utilized an efficient channel attention and self-calibrated convolutions (ECA-SC) module for feature enhancement.
• Introduced distribution ranking (DR) loss to address class imbalance in positive and negative samples.

Main Results:

• MSANet achieved a Competitive Detection Performance (CPM) score of 0.920 on the LUNA16 dataset.
• Demonstrated improved sensitivity in detecting pulmonary nodules.
• Significantly reduced the average number of false positives compared to other networks.

Conclusions:

• MSANet offers superior performance in 3D pulmonary nodule detection.
• The proposed network effectively assists radiologists in diagnosing lung cancer.
• MSANet shows significant potential for clinical application in early lung cancer detection.