Lightweight Lung-nodule Detection Model Combined with Multidimensional Attention Convolution.

Area of Science:

• Artificial Intelligence in Medical Imaging
• Deep Learning for Cancer Detection
• Computer-Aided Diagnosis Systems

Background:

• Early detection of pulmonary nodules is critical for improving lung carcinoma survival rates.
• Current Convolutional Neural Network (CNN) models struggle with low accuracy and detecting small nodules.
• Existing accurate CNN models are computationally expensive and require high hardware specifications.

Purpose of the Study:

• To develop a novel detection model for pulmonary nodules.
• To achieve both high accuracy and real-time performance in nodule detection.
• To facilitate effective and timely diagnosis of lung carcinoma.

Main Methods:

• Modified YOLOv5s architecture incorporating a C3_ODC module with multidimensional attention for enhanced feature extraction.
• Developed a GS-BiFPN structure using lightweight convolution and weighted BiFPNs for improved multiscale feature fusion and parameter reduction.
• Optimized the loss function using Focal Loss to address class imbalance and Normalized Wasserstein Distance (NWD) for enhanced small nodule detection.

Main Results:

• The proposed model demonstrated an 8.7% improvement in average precision compared to YOLOv5s.
• Achieved a reduction of 5.4% in model parameters and 8.2% in floating-point operations.
• Real-time performance was achieved with a processing speed of 116.7 frames per second.

Conclusions:

• The developed model effectively balances high detection accuracy with real-time processing requirements.
• This advancement offers a promising solution for timely and accurate pulmonary nodule detection in clinical settings.