Conditional GAN with 3D discriminator for MRI generation of Alzheimer's disease progression

Abstract

Many studies aim to predict the degree of deformation on affected brain regions as Alzheimer's disease (AD) progresses. However, those studies have been often limited since it is difficult to obtain sequential longitudinal MR data of affected patients. Recently, conditional generative adversarial networks (cGANs) have been used to estimate the changes between unpaired images by modeling their differences. However, generating high-quality 3D magnetic resonance (MR) brain images with cGANs requires a large amount of computation. Previous models have been mostly designed to operate in 2D space taking individual slices or down-sampled 3D space, but these approaches often cause spatial artifacts such as discontinuities between slices or unnatural changes in 3D space. To address these limitations, we propose a novel cGAN that can synthesize high-quality 3D MR images at different stages of AD by integrating an additional module that ensures smooth and realistic transitions in 3D space. Specifically, the proposed cGAN model consists of an attention-based 2D generator, a 2D discriminator, and a 3D discriminator that is able to synthesize continuous 2D slices along the axial view resulting in good quality 3D MR volumes. Moreover, we propose an adaptive identity loss so that relevant transformations take place without compromising the features to identify patients. In our experiments, the proposed method showed better image generation performance than previously proposed GAN methods in terms of image quality and image generation suitable for the condition. © 2022 Elsevier Ltd