Channel-Wise Attention Mechanism in the 3D Convolutional Network for Lung Nodule Detection

Pulmonary nodule detection is essential to reduce the mortality of lung cancer. One-stage methods have recently emerged as high-performance and lower-power alternatives two-stage methods. However, it difficult for existing one-stage networks balance sensitivity specificity. In this paper, we propose an end-to-end mechanism combined with a channel-wise attention based on 3D U-shaped residual network. First, improved gate (AG) introduced false positive rate by employing critical feature dimensions at skip connections propagation. Second, channel interaction unit (CIU) designed before head further improve sensitivity. Furthermore, gradient harmonizing (GHM) loss function adopted solve problem caused imbalance negative samples. We conducted experiments LUNA16 dataset achieved performance competition metric (CPM) score 89.5% 95%. The proposed method outperforms models in terms specificity while maintaining attractiveness being lightweight, making suitable automatic detection.