A multi-scale self-calibrating lung nodule detection based on SPC-UNet

Early-stage lung cancer usually manifests itself in the form of lung nodules, so accurate detection and early screening of lung nodules based on CT images is key to early lung cancer treatment and reduction of the risk of mortality. However, the automatic detection of small lung nodules is still a challenge because shape, tissue contrast and other factors affect the detection. In order to improve the rate and accuracy of lung nodule detection, this paper proposes SPC-UNet, a Res U-Net lung nodule detection algorithm based a self-calibrated multi-scale attention mechanism. In order to target lung nodules with different sizes and structures and extract the corresponding scale features in the multi-scale dimension, Self-calibrated Pyramidal Squeeze Attention module (SC-PSA module) is designed in this paper. The Coordinate Attention module (CA module) and SCPSA module in SPC-UNet are used to extract reliable discriminative information between features from both spatial and channel perspectives, which is beneficial for benign and malignant lung nodule detection. On the representative Luna16 dataset, the proposed SPC-UNet network achieves 96.18% accuracy and 0.924 CPM in lung nodule detection performance.