The lightweight architecture and anatomical context-based lung nodule detection neural network

With the increasing demands for computational power and storage resources in primary healthcare, mobile terminals, and edge computing scenarios, existing lung nodule detection models often struggle to achieve sufficient lightweight optimization while maintaining high detection accuracy, which limits their practical clinical application. To address this issue, this paper proposes a lightweight lung nodule detection method based on YOLOv8. First, an improved Ghost convolution module is designed to perform hierarchical feature extraction and redundancy reduction, effectively reducing the number of convolution kernels and floating-point operations, thus enhancing the inference efficiency of the model on low-power platforms. Secondly, anatomical and pathological prior knowledge of the lungs is incorporated into the feature fusion stage, using heatmap-guided mechanisms to strengthen the network’s focus on suspected nodule regions. Meanwhile, depthwise separable convolutions are used to replace standard convolutions, further compressing the model size while preserving feature representation capabilities. Finally, experimental results on the LUNA16 public dataset show that the proposed method significantly reduces model parameters and computational costs while maintaining detection accuracy and recall rates comparable to the original YOLOv8. This approach strikes a balance between performance and efficiency, offering a feasible solution for deploying lung nodule detection models in resource-constrained environments, with promising clinical application prospects.