Improved cortical surface reconstruction using sub-millimeter resolution MPRAGE by image denoising

Abstract

Automatic cerebral cortical surface reconstruction is a useful tool for cortical anatomy quantification, analysis and visualization. Recently, the Human Connectome Project and several studies have shown the advantages of using T1-weighted magnetic resonance (MR) images with sub-millimeter isotropic spatial resolution instead of the standard 1-millimeter isotropic resolution for improved accuracy of cortical surface positioning and thickness estimation. Nonetheless, sub-millimeter resolution images are noisy by nature and require averaging multiple repetitions to increase the signal-to-noise ratio for precisely delineating the cortical boundary. The prolonged acquisition time and potential motion artifacts pose significant barriers to the wide adoption of cortical surface reconstruction at sub-millimeter resolution for a broad range of neuroscientific and clinical applications. We address this challenge by evaluating the cortical surface reconstruction resulting from denoised single-repetition sub-millimeter T1-weighted images. We systematically characterized the effects of image denoising on empirical data acquired at 0.6 mm isotropic resolution using three classical denoising methods, including denoising convolutional neural network (DnCNN), block-matching and 4-dimensional filtering (BM4D) and adaptive optimized non-local means (AONLM). The denoised single-repetition images were found to be highly similar to 6-repetition averaged images, with a low whole-brain averaged mean absolute difference of ∼0.016, high whole-brain averaged peak signal-to-noise ratio of ∼33.5 dB and structural similarity index of 0.92, and minimal gray matter–white matter contrast loss (2% to 9%). The whole-brain mean absolute discrepancies in gray–white surface placement, gray–CSF surface placement and cortical thickness estimation were lower than 165 μm, 155 μm and 145 μm—sufficiently accurate for most applications. The denoising performance is equivalent to averaging ∼2.5 repetitions of the data in terms of image similarity, and 1.6–2.2 repetitions in terms of the cortical surface placement accuracy. The scan-rescan precision of the cortical surface positioning and thickness estimation was lower than 170 μm. Our unique dataset and systematic characterization support the use of denoising methods for improved cortical surface reconstruction sub-millimeter resolution.

High-fidelity fast volumetric brain MRI using synergistic wave-controlled aliasing in parallel imaging and a hybrid denoising generative adversarial network

“…Even the simplest CNN for denoising (i.e., DnCNN 17 ) achieves superior performance compared to the state-of-the-art block-matching and 3D filtering (BM3D) 21 denoising method. Recently, it has been shown that DnCNN is also more advantageous in removing noise from sub-millimeter resolution T 1 -weighted structural MR image volumes 22 compared to the state-of-the-art block matching with 4D filtering (BM4D) 23 and adaptive optimized nonlocal means (AONLM) 24 . However, a well-known problem of CNNs trained using the voxel-wise error as the loss function is the tendency to generate blurry images that lack realistic textural details in both super-resolution and denoising tasks, even though the mean squared error (MSE), mean absolute error (MAE) or related metrics such as peak SNR (PSNR) can be minimized or maximized 25–27 , respectively.…”

“…The advantage of 2D GANs is that their parameters can be optimized on only a few subjects, because each image volume from a subject provides millions of voxels as training samples for calculating the voxel-wise loss for the generator and hundreds of image slices as training samples for calculating the image-wise loss for the discriminator. However, the image synthesis performance of 2D generators is often limited compared to 3D generators, which can incorporate complementary information from an additional spatial dimension 22,38,39 (Supplementary Information Figure 1). Moreover, there may be boundary artifacts across synthesized 2D image slices along the cross-sectional direction.…”

“…Moreover, there may be boundary artifacts across synthesized 2D image slices along the cross-sectional direction. Consequently, 3D CNNs are more often adopted for volumetric data 20,22,38–40 and 3D GANs consisting of a 3D generator and a 3D discriminator have been also proposed 34–36 .…”