Generating Synthetic Brain Tumor Data Using StyleGAN3 for Lower Class Enhancement

Accurate simulation of brain tumor data is crucial in medical imaging research for developing and assessing novel image enhancement techniques. We propose an approach to generate synthetic brain tumor data using StyleGAN3, a GAN model. Our objective is to explore the impact of varying gamma values and dataset sizes on the quality of synthetic brain tumor images via transfer learning from a pre-trained 512 X 512 model. We conducted three experiments with different gamma values and dataset sizes. The first two experiments employed gamma values of 6.6 and 8.2, respectively, using a dataset of 310 brain tumor images. The third experiment maintained the gamma value at 8.2 but reduced the dataset size to 155 images. We used FID and KID as evaluation metrics. Results show varying gamma values and dataset sizes significantly influence synthetic image quality. Higher gamma values generally yielded improved image quality. In another experiment, we obtained FID and KID scores of 67.856 and 0.013, respectively [1]. These findings impact brain tumor imaging enhancement techniques. StyleGAN3 and transfer learning can efficiently generate synthetic brain tumor datasets, advancing medical image analysis and diagnosis. Our study underscores synthetic data generation’s potential for enhancing medical image research.