Objective To propose a method based on deep belief network (DBN) to automatically identify pulmonary nodules so as to improve the detection accuracy of pulmonary nodules.Methods To meet the training sample requirements of DBN,a database of 4 000 lung nodule images identified by professional doctors was established,and the sample database was expanded using virtual sample technology.In this technology,new samples of the database were generated from the manually recognized region of interest (ROI) by rotation,scaling and panning,or by a series of combinations of two or more operations of panning,scaling,rotation,and compositing.Finally,some samples from the sample database were input into the convolutional neural network classifier,and the ROI of the suspected pulmonary nodule was output by optimizing the network parameters.Result The sample size of the training sample database was expanded to 40 000 using the virtual sample expansion.Based on the training database obtained by this method,the detection accuracy of DBN for identifying pulmonary nodules was 90%,and the false positive rate was 0.4%.Conclusion Virtual sample technology can effectively improve the efficiency of training database establishment.The accuracy of using DBN-based CAD technology to detect pulmonary nodules is high,allowing doctors to focus only on areas where lung nodules are detected,thus effectively improving the efficiency of diagnosis.