Because CT lung nodules vary in size, shape, and texture, it is extremely difficult to diagnose benign and malignant lung nodules. Based on three-dimensional convolutional neural network, a network based on multi-depth residual attention mechanism (MDRA-net) is proposed to classify benign and malignant pulmonary nodules. The MDRA-net improves the network's perception of nodule location and global features using feature fusion and iterative hierarchical fusion on residual differential branches. Furthermore, combined with the attention mechanism, the projection and excitation block module is introduced to calibrate with spatial and channel information, which can further improve the ability of the network to extract features. Experimental results on the LUNA16 dataset show that the accuracy of the MDRA-net classification model is 96.52%, and the sensitivity and specificity are 93.01% and 97.77%, respectively, which are greatly improved compared with those of the existing classification methods of lung nodules, based on deep learning.