3d brain mri gan-based synthesis conditioned on partial volume maps

Abstract

In this paper, we propose a framework for synthesising 3D brain T1-weighted (T1-w) MRI images from Partial Volume (PV) maps for the purpose of generating synthetic MRI volumes with more accurate tissue borders. Synthetic MRIs are required to enlarge and enrich very limited data sets available for training of brain segmentation and related models. In comparison to current state-of-the-art methods, our framework exploits PV-map properties in order to guide a Generative Adversarial Network (GAN) towards the generation of more accurate and realistic synthetic MRI volumes. We demonstrate that conditioning a GAN on PV-maps instead of Binary-maps results in 58.96% more accurate tissue borders in synthetic MRIs. Furthermore, our results indicate an improvement in the representation of the Deep Gray Matter region in synthetic MRI volumes. Finally, we show that fine changes introduced into PV-maps are reflected in the synthetic images, while preserving accurate tissue borders, thus enabling better control during the data synthesis of novel synthetic MRI volumes.