Chen Junlong

School of Economics and Management, Jiangxi Normal University

Abstract：

Lung cancer is the leading cause of cancer mortality worldwide. Although CT is the primary screening modality, manual interpretation suffers from slow speed and high miss rates. This paper integrates dynamic fusion generative adversarial networks (GANs) with multi-scale attention mechanisms to enhance YOLOv10 for pulmonary micronodule detection, addressing limitations in feature extraction for micro-size lesions and high false-positive rates. Experimental results on the LUNA16 dataset demonstrate a 3.5 percentage point improvement in CPM and a 30% reduction in false-positive rate, while maintaining 55 FPS. Compared with baseline models, our approach achieves notable improvement in micronodule detection.

Key Words：

channel-spatial attention mechanism; generative adversarial networks; pulmonary micronodule detection; YOLOv10