Lung nodule detection of CT images based on combining 3D-CNN and squeeze-and-excitation networks

Malignant lung nodules are the worse stage for lung cancer patients. Early detection of lung nodules is essential for early treatment of the patients and significantly improves their survival rate. Recently, several computer-aided diagnosis (CAD) schemes based on deep learning approaches have been suggested to assist in lung nodule diagnosis. This study presents a 3D U-shaped encoding and decoding deep convolutional neural network (CNN) integrated with channel attention mechanisms for lung nodule detection in chest CT images. The U-shaped network applies the encoder to extract nodule representation features and the decoder to indicate the prediction results of the nodule candidates. Hybrid efficient channel attention (ECA) modules are integrated to enhance network representation power by retaining rich contextual nodule information and suppressing useless features. Correspondingly, the 3D regional proposal network (RPN) with three anchor boxes is employed for multilevel nodule candidates detection. The Lung Nodule Analysis 2016 (LUNA16) dataset was used to validate the proposed method. Experimental results show that through 10-fold cross-validation, the proposed algorithm gives the highest detection sensitivity of 98.65% and the competition performance metric (CPM) value of 0.911. Therefore, the results of our design show competitiveness compared with the methods in the prior studies for lung nodule detection in chest CT images.