Lung cancer is the deadliest diseases, and adenocarcinoma in them accounts for nearly 40% due to high mortality and morbidity. Most patients suffer from the advanced lung cancer rather than the early one which can be cured by medical intervention. Different from the advanced lung cancer, the early lung adenocarcinoma is shown as ground glass nodules (GGNs) in CT images. GGNs are insidious, symptoms, and inconspicuous, making patient be misdiagnosed. In this paper, considering the powerful feature learning capability of convolutional neural network (CNN) and the low similarity between medical images and natural images, we proposed a benign and malignant classification method after improving the existing feature-based transfer learning method, where the use of transfer learning method aims at overcoming the problem of deficient GGN samples. Essentially, we extracted the features of GGNs by determining the network layer one by one from different pre-trained CNN networks. Then the optimal layer with the highest classification accuracy is found compared with the existing feature-based transfer learning method. When transferring the features from the "pool5" layer of AlexNet, the classification accuracy can attain 95.3%, the sensitivity 97.1%, and the specificity 93.3%, respectively. So far, the classification result is best compared with other results in literature.