An Enhanced 3D YOLOv5 for Pulmonary Nodule Detection Based on Backbone and Neck Optimization

Accurate detection of pulmonary nodules is critical for early lung cancer screening. However, existing 3D YOLOv5-based detection networks suffer from redundant feature extraction and insufficient contextual modeling. This paper proposes an optimized 3D YOLOv5 network for pulmonary nodule detection, with two core improvements to enhance performance: (1) a Cross-Stage Partial Single Aggregation (CSP_OSA) module, which reduces memory consumption by 30% through feature compression while enhancing multi-scale feature representation; and (2) a Dual Attention and Multi-Scale Receptive Field Fusion (DA_DRFF) module, which dynamically filters key features and integrates contextual information from surrounding tissues to suppress noise interference. Experiments on the LUNA16 dataset demonstrate that the improved model achieves a sensitivity of 96.8% at a false positive rate (FPs/scan) of 8, with a comprehensive performance metric (CPM) of 0.898, significantly outperforming mainstream models such as DeepLung and S4ND. Generalization experiments further validate the robustness of the model (sensitivity of 96.3% at FPs/scan = 8). This study provides an efficient and reliable 3D medical imaging analysis framework for early lung cancer screening, with significant clinical application value.