Dynamic Convolution Enhanced Attention Network for Pulmonary Nodule Detection

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Pulmonary nodules on CT scans are crucial for lung cancer screening.
• Existing lightweight deep learning models struggle with accuracy for small nodules and have redundant parameters.

Purpose of the Study:

• To develop an improved, lightweight deep learning model for enhanced pulmonary nodule detection.
• To address limitations in accuracy for tiny lesions and network redundancy in current models.

Main Methods:

• An enhanced YOLOv8n model incorporating Omni-Dimensional Dynamic Convolution (ODConv) for feature extraction.
• Integration of Convolutional Block Attention Module (CBAM) to reduce background interference.
• Utilizing Wise Intersection over Union (W-IoU) loss for optimized bounding box regression.

Main Results:

• The proposed model demonstrated improved Precision (6.3%), Recall (8.6%), mAP50 (3.4%), and mAP50-95% (2.7%) on the LUNA16 dataset.
• The model maintained high inference speed compared to the original YOLOv8n.
• Validation on the LIDC-IDRI dataset confirmed the model's robustness and balanced performance.

Conclusions:

• The improved YOLOv8n architecture effectively enhances pulmonary nodule detection accuracy, especially for small lesions.
• The model offers a robust, lightweight solution balancing accuracy and real-time performance for lung cancer screening.