Synthetic generation and evaluation of brain MRI images with generative adversarial networks

The number and quality of data are very important in artificial intelligence studies with medical images. Especially the failure to obtain ethical permissions interrupts the studies. The main purpose of this study is to provide a solution to the data problem in health studies. Using DCGAN, this problem is avoided by producing synthetic images that do not require ethical permission. The quality of the synthetic images is evaluated by measuring the Frechet Inception Distance (FID) and Inception Score (IS). The aim of this study is to investigate the effect of increasing the training time and the number of real data on the quality of the synthetic images produced. The study used datasets consisting of brain tumor MRI images from open source. The datasets used consist of 3 classes: Glioma tumor, pituitary tumor, and no tumor. The study consists of two parts: with little data and with a lot of data. In both studies, the classes were trained with different epoch training times and FID and IS scores were calculated for each training. As a result, it is seen that increasing the number of data and training time positively affects the quality of synthetic images produced with DCGAN. It is predicted that synthetic data can be used as a solution to the data problem in the health field.