Abdominal MRI Synthesis using StyleGAN2-ADA

Abstract

The lack of labelled medical data still poses as one of the biggest issues when creating Deep Learning models in the medical field. Modern data augmentation techniques like the generation of synthetic images have gained a special interest. In recent years there has been a significant improvement in GANs. StyleGAN2 achieves impressive results in the generation of natural images. StyleGAN2-ADA was created to respond to the lack of training data when training an image synthesis model, which is very frequent in the medical field. Some works used styleGAN to generate melanomas, breast cancer histological images, MR and CT images. In this work we apply, for the first time, a styleGAN2-ADA to a small dataset of abdominal MRI with 1.3k images. From the augmentation pipeline created by the authors of styleGAN2-ADA, we removed all augmentations except the geometric transformations and pixel blitting operations. We trained our network for 70 hours. Our generated dataset has a precision score of 59,33 % and a FID score of 18,14. We conclude that the styleGAN2-ADA is a viable solution to generate MRI using a small dataset.