I am currently working on training a Generative Adversarial Network (GAN) model and I am facing some challenges. I have a dataset of approximately 100 MRI slices of the pelvis, but it seems that this amount of data is not sufficient to train a robust GAN model. The results I am getting from the GAN are not satisfactory.
My question is twofold:
What is the typical number of images required to train a GAN model effectively? Are there any recommended data augmentation methods that could help me expand my dataset or potentially improve the performance of my GAN model? Any advice or guidance would be greatly appreciated. Thank you in advance!
Regarding the question, "What is the typical number of images required to train a GAN model effectively?" Actually, it is difficult to pinpoint a specific number that guarantees effective training. The model's performance depends on various factors such as data diversity, color, shape, quality, and more. Many GAN implementations use datasets like ImageNet or CIFAR, which contain a wide range of diverse images encompassing animals, planets, cars, and more. However, in the case of medical images, such as MRIs, the variations between different data are subtle, and the images are often in grayscale.
I recommend referring to the paper titled "Training Generative Adversarial Networks with Limited Data." This paper addresses the issue of poor GAN performance, primarily attributed to discriminator overfitting. The authors propose the Adaptive Discriminator Augmentation method to tackle this problem. Their experiments also include testing on medical image datasets like BRECAHAD.
Paper Link: https://arxiv.org/pdf/2006.06676.pdf