In view of the low accuracy and high false positive of automatic detection model of pulmonary nodules, a two-stage detection method of pulmonary nodules has thus been proposed based on 3D convolution neural network. In the initial stage, 3D V-Net is used for the detection of all candidate nodules, with the residual jump connection integrated to build a deep network, so as to retain a certain proportion of output from the upper network, thus realizing image feature reuse, followed by an introduction of an improved loss function to solve the problem of imbalance between positive and negative samples in the data set. In the second stage, 3D VGG network is used to classify candidate nodules to reduce false positives, with the residual connection added to prevent gradient loss and degradation, thus accelerating the network training process. Experiments show that the proposed method achieves a sensitivity rate as high as 91.28% in candidate nodule detection stage, an accuracy rate of 99.22% and a sensitivity rate of 96.60% in classification stage respectively, which can effectively assist radiologists in the detection of pulmonary nodules.