Despite the rapidly growing interest, progress in the study of relations between physiological abnormalities and mental disorders is hampered by complexity of the human brain and high costs of data collection. The complexity can be captured by deep learning approaches, but they still may require significant amounts of data. In this paper, we seek to mitigate the latter challenge by developing a generator for synthetic realistic training data. Our method greatly improves generalization in classification of schizophrenia patients and healthy controls from their structural magnetic resonance images. A feed forward neural network trained exclusively on continuously generated synthetic data produces the best area under the curve compared to classifiers trained on real data alone.