High-fidelity fast volumetric brain MRI using a hybrid denoising generative adversarial network (HDnGAN)

PurposeTo improve the signal-to-noise ratio (SNR) of highly accelerated volumetric MRI while preserve realistic textures using a generative adversarial network (GAN). MethodsA hybrid GAN for denoising entitled "HDnGAN" with a 3D generator and a 2D discriminator was proposed to denoise 3D T2-weighted fluid-attenuated inversion recovery (FLAIR) images acquired in 2.75 minutes (R=3x2) using wave-controlled aliasing in parallel imaging (Wave-CAIPI). HDnGAN was trained on data from 25 multiple sclerosis patients by minimizing a combined mean squared error and adversarial loss with adjustable weight{lambda} . Results were evaluated on eight separate patients by comparing to standard T2-SPACE FLAIR images acquired in 7.25 minutes (R=2x2) using mean absolute error (MAE), peak SNR (PSNR), structural similarity index (SSIM), and VGG perceptual loss, and by two neuroradiologists using a five-point score regarding gray-white matter contrast, sharpness, SNR, lesion conspicuity, and overall quality. ResultsHDnGAN ({lambda}=0) produced the lowest MAE, highest PSNR and SSIM. HDnGAN ({lambda}=10-3) produced the lowest VGG loss. In the reader study, HDnGAN ({lambda}=10-3) significantly improved the gray-white contrast and SNR of Wave-CAIPI images, and outperformed BM4D and HDnGAN ({lambda}=0) regarding image sharpness. The overall quality score from HDnGAN ({lambda}=10-3) was significantly higher than those from Wave-CAIPI, BM4D, and HDnGAN ({lambda}=0), with no significant difference compared to standard images. ConclusionHDnGAN concurrently benefits from improved image synthesis performance of 3D convolution and increased training samples for training the 2D discriminator on limited data. HDnGAN generates images with high SNR and realistic textures, similar to those acquired in longer times and preferred by neuroradiologists.