: In the pulmonary nodules computer aided diagnosis systems (CAD, feature selection plays an important role in reducing the false positive rate and improving the system accuracy. To solve the problem of feature selection techniques by which the diversity of featureswas damaged in the process of distinguishing malignant pulmonary nodules from benign pulmonary nodules, this study developed a novel feature selection algorithm for improving the accuracy of traditional computer-aided differential diagnosis for benign and malignant classification of pulmonarynodules.;: Firstly, we divided the extracted features of nodules into several groups by using Gaussian mixture model (GMM. Secondly, we applied Relief and sequential forward selection (SFS algorithm to find local optimum features dataset... for each group. Afterwards, we used theoptimumpath forest (OPF classifier with the found features dataset to obtain the classification results. Finally, the local optimum features dataset with the highest area under curve AUC in all groups were added into the final selected set.;: According to collected pulmonarynodules on computed tomography (CT scans, tested with two set of samples, we achieved an average accuracy of 89.5%, sensitivity of 87.1% and specificity of 90.9% on the first set of samples, and 90.1%, 88.7% and 92.1% on the second set of samples. The areas under the receiver operating characteristic(ROC curves based on these two sample sets were 95.2%, and 96.3% respectively.;: This study shows that the proposed method was promising for improving the pulmonary nodules computer aided diagnosis systems performance of benign and malignant pulmonary nodules.