Synthesis of Wilms tumor images using Generative Adversarial Networks

Wilms tumor or Nephroblastoma is paediatric cancer, that can be treated if diagnosed at an early stage. However, no appropriate image dataset is publicly available except the Wilms MRI dataset of Mathematical Image Analysis Group, Saarland University, Germany, which only contain 28 multisequence MR scans. As, huge amount of data is required to train a machine or deep learning model, image synthesis could be beneficial. The absence of professionals who can label data was another problem. In order to meet the data requirements two different Generative Adversarial Networks (GANs) - Wasserstein GAN (WGAN) and Pix2Pix GAN have been utilized. These GAN architectures generate high-quality and realistic MRI images in terms critical evaluation metrics. These architectures may potentially benefit the detection of rare diseases which lack sufficient size of medical datasets. With the help of image synthesis and precise amount of data, effective computer-aided diagnostic models could be developed.