Deep learning methods have significantly improved classification accuracy in different areas such as speech, object and text recognition. However, this field has only began to be explored in the brain imaging field, which differs from other fields in terms of the amount of data available, its data dimensionality and other factors. This paper proposes a methodology to generate an extensive synthetic structural magnetic resonance imaging (sMRI) dataset to be used at the pre-training stage of a shallow network model to address the issue of having a limited amount of data available. Our results show that by extending our dataset using 5,000 synthetic sMRI volumes for pretraining, which accounts to approximately 10 times the size of the original dataset, we can obtain a 5% average improvement on classification results compared to the regular approach on a schizophrenia dataset. While the use of synthetic sMRI data for pre-training has only been tested on a shallow network, this can be readily applied to deeper networks.