High-fidelity fast volumetric brain MRI using a hybrid denoising generative adversarial network (HDnGAN)
PurposeReducing scan times is important for wider adoption of high-resolution volumetric MRI in research and clinical practice. Emerging fast imaging and deep learning techniques provide promising strategies to accelerate volumetric MRI without compromising image quality. In this study, we aim to leverage an advanced fast imaging technique, wave-controlled aliasing in parallel imaging (Wave-CAIPI), and a novel denoising generative adversarial network (GAN) to achieve accelerated high-fidelity, high-signal-to-noise-ratio (SNR) volumetric MRI. Methods3D T2-weighted fluid-attenuated inversion recovery (FLAIR) image data were acquired on 33 multiple sclerosis (MS) patients using a prototype Wave-CAIPI sequence (acceleration factor R=3x2, 2.75 minutes) and a standard T2-SPACE FLAIR sequence (R=2, 7.25 minutes). A hybrid denoising GAN entitled "HDnGAN" composed of a 3D generator (i.e., a modified 3D U-Net entitled MU-Net) and a 2D discriminator was proposed to denoise Wave-CAIPI images with the standard FLAIR images as the target. HDnGAN was trained and validated on data from 25 MS patients by minimizing a combined content loss (i.e., mean squared error (MSE)) and adversarial loss with adjustable weight{lambda} , and evaluated on data from 8 patients unseen during training. The quality of HDnGAN-denoised images was compared to those from other denoising methods including AONLM, BM4D, MU-Net, and 3D GAN in terms of their similarity to standard FLAIR images, quantified using MSE and VGG perceptual loss. The images from different methods were assessed by two neuroradiologists using a five-point score regarding sharpness, SNR, lesion conspicuity, and overall quality. Finally, the performance of these denoising methods was compared at higher noise levels using simulated data with added Rician noise. ResultsHDnGAN effectively denoised noisy Wave-CAIPI images with sharpness and rich textural details, which could be adjusted by controlling{lambda} . Quantitatively, HDnGAN ({lambda}=10-3) achieved low MSE (7.43 x10-4{+/-}0.94x10-4) and the lowest VGG perceptual loss (1.09x10-2{+/-}0.18x10-2). The reader study showed that HDnGAN ({lambda}=10-3) significantly improved the SNR of Wave-CAIPI images (4.19{+/-}0.39 vs. 2.94{+/-}0.24, P<0.001), outperformed AONLM (4.25{+/-}0.56 vs. 3.75{+/-}0.90, P=0.015), BM4D (3.31{+/-}0.46, P<0.001), MU-Net (3.13{+/-}0.99, P<0.001) and 3D GAN ({lambda}=10-3) (3.31{+/-}0.46, P<0.001) regarding image sharpness, and outperformed MU-Net (4.21{+/-}0.67 vs. 3.29{+/-}1.28, P<0.001) and 3D GAN ({lambda}=10-3) (3.5{+/-}0.82, P=0.001) regarding lesion conspicuity. The overall quality score of HDnGAN ({lambda}=10-3) (4.25{+/-}0.43) was significantly higher than those from Wave-CAIPI (3.69{+/-}0.46, P=0.003), BM4D (3.50{+/-}0.71, P=0.001), MU-Net (3.25{+/-}0.75, P<0.001), and 3D GAN ({lambda}=10-3) (3.50{+/-}0.50, P<0.001), with no significant difference compared to standard FLAIR images (4.38{+/-}0.48, P=0.333). The advantages of HDnGAN over other methods were more obvious at higher noise levels. ConclusionHDnGAN provides robust and feasible denoising while preserving rich textural detail in empirical volumetric MRI data and is superior on both quantitative and qualitative evaluation compared to the original Wave-CAIPI images and images denoised using standard methods. HDnGAN concurrently benefits from the improved image synthesis performance of the 3D convolution and the increased number of samples for training the 2D discriminator from a limited number of subjects. Our study supports the use of HDnGAN in combination with modern fast imaging techniques such as Wave-CAIPI to achieve high-fidelity fast volumetric MRI.