Attention-Guided Feature Extraction and Multiscale Feature Fusion 3D ResNet for Automated Pulmonary Nodule Detection
Guanglu Zhang, Hongjun Zhang, Yuhua Yao, Qiuhui Shen
Abstract
Automatic detection of pulmonary nodules is of great significance for early diagnosis and prevention of lung cancer. Computed tomography (CT) is an effective and economical detection method. In CT images, the size and shape of pulmonary nodules are different, and some nodules are very similar to the surrounding tissues. Therefore, the automatic localization of pulmonary nodules in CT images is a challenging task. In this study, an attention embedded three-dimensional convolutional neural network is proposed for pulmonary nodule detection. Specifically, 1) channel-spatial attention guides 3D ResNet to down sample the input 3D CT patch. The channel pays attention to important features and the space pays attention to the region of interest. The two form a complementary feature extraction mechanism to effectively help the global flow of information in the network, and further refine the feature mapping to extract the nodule context features. 2) The channel spatial attention module changes the fusion mode of feature pyramid, adaptively adjusts the pixel level weight between features, and extracts multi-scale representative node features. 3) The deep separable convolution is used to replace the standard convolution of ResNet, which can reduce the time cost and improve the efficiency of model training on the premise of ensuring the performance of the model. 4) In order to adapt to the distribution of nodule scale, different characteristic layers correspond to two sizes of anchors. Under the condition of ensuring the detection rate of nodules, the number of anchor frames is reduced and the network sensitivity is improved. Finally, a large number of ablation experiments are carried out in the on the LUNA16 dataset. The results show that the attention guided network can extract the multi-scale representative features of nodules, and the average sensitivity is 97.7%. At the same time, the CMP score reached 0.912. The detection effect of this method is better than most detection methods based on deep learning, and has a certain reference value.