Graphs provide a way to study the characteristics of structural and functional connections between different brain regions. Graph convolutional neural networks have the ability to extract intrinsic local characteristics of networks. However, lack of data is a common problem in medical field, holding back the application of deep learning. Although transfer learning is an effective method to improve performance, it is difficult to find natural graph-structured datasets to pre-train deep learning models. To address this problem, we proposed a novel transfer learning method. The method uses grid-structured source data to pre-train a model, and fine-tuned it with graph-structured data in the task of interest. At last, the method is applied to the diagnosis of depression with a total of 83 samples. By comparing the performance between a fine-tuned model and a fully trained model, we tested the effectiveness of the proposal method. And there is a significant improvement in the accuracy of the pre-trained model.