Integrated 3D Deep Learning Approach: YOLO-Decoupled Candidate Detection and 3D CNN-Driven False Positive Reduction for Lung Nodule Analysis in CT

Lung cancer often presents with small pulmonary nodules in its early stages, making early detection and diagnosis of these nodules vital for treatment. This paper proposes a two-step automatic pulmonary nodule detection method. For candidate nodule detection, we decouple detection and classification in YOLO for axial slice analysis, using an Attention Feature Fusion ResBlock (AFF-ResBlock) to enhance the network and weighted cluster non-maximum suppression (NMS) for post-processing. In false positive reduction, a 3D CNN architecture is designed. 3D CNNs capture richer 3D spatial information than 2D ones. We also integrate multi-scale contextual information. Our algorithm achieves 95.3\% accuracy on the LUNA16 dataset, proving its effectiveness.