Background: A generative adversarial network (GAN) has gained popularity as a data augmentation technique in the medical field due to its efficiency in creating synthetic data for different machine learning models. In particular, the earlier literature suggests that the classification accuracy of a convolutional neural network (CNN) used for detecting brain tumors in magnetic resonance imaging (MRI) images increases when GAN-generated images are included in the training data together with the original images. However, there is little research about how the exact number of GAN-generated images and their ratio to the original images affects the results obtained. Materials and methods: Here, by using 1000 original images from a public repository with MRI images of patients with or without brain tumors, we built a GAN model to create synthetic brain MRI images. A modified U-Net CNN is trained multiple times with different training datasets and its classification accuracy is evaluated from a separate test set of another 1000 images. The Mann-Whitney U test is used to estimate whether the differences in the accuracy caused by different choices of training data are statistically significant.

Results: According to our results, the use of GAN augmentation only sometimes produces a significant improvement. For instance, the classification accuracy significantly increases when 250-750 GAN-generated images are added to 1000 original images (p-values ≤ 0.0025) but decreases when 10 GAN-generated images are added to 500 original images (p-value: 0.03). Conclusions: Whenever GAN-based augmentation is used, the number of GAN-generated images should be carefully considered while accounting for the number of original images.