Abstract:
Machine Learning (ML) is widely utilized in various fields to solve problems and assist
research. In the field of medical image synthesis, Deep Learning (DL) shows great
potential. Recently, different kinds of frameworks of Generative Adversarial Network
(GAN) have been developed. However, the reliability of the synthetic images cannot be
guaranteed as standard GANs only produce point estimates. In this situation, uncertainty
quantification is of crucial importance to evaluate the confidence of the model in its
prediction. In this work, we perform the image synthesis based on the two Magnetic
Resonance (MR) datasets of different pathologies. The baseline network for the Multiple
Sclerosis (MS) dataset is CycleGAN, while pix2pix is applied for the Glioma dataset.
In order to model the predictive uncertainty, two methods using Bayesian techniques
are employed: the Monte Carlo (MC) dropout and the Bayesian GAN using Variational
Inference (VI). Moreover, a comparison study is conducted to show the advantages and
disadvantages of the two presented methods. Moreover, suggestions for their improvement
are discussed.