Transformed domain NORDIC (tNORDIC) denoising improves mesoscopic, whole brain quantitative imaging

2299 

Abstract Submission 

Logan Dowdle¹, Steen Moeller¹, Omer Faruk Gulban², Luca Vizioli¹, Kamil Uğurbil¹

¹Center for Magnetic Resonance Research, Minneapolis, MN, ²Maastricht University, Maastricht, Netherlands

Logan Dowdle, Ph.D.  
Center for Magnetic Resonance Research
Minneapolis, MN

Steen Moeller, Ph.D.  
Center for Magnetic Resonance Research
Minneapolis, MN

Omer Faruk Gulban, Ph.D.  
Maastricht University
Maastricht, Netherlands

Luca Vizioli, Ph.D.  
Center for Magnetic Resonance Research
Minneapolis, MN

Kamil Uğurbil, Ph.D.  
Center for Magnetic Resonance Research
Minneapolis, MN

With the advent of MRI scanners with field strengths of 7T and above there is an increased interest in ultra-high high-resolution anatomical images (<0.5mm) of living human brains. These images have many uses, including structural references for depth-dependent fMRI, biophysical modeling of the BOLD signal, and uncovering tissue properties. Unfortunately, the acquisition of these images is associated with several challenges, including long scan times, and participant motion. Ideally, we could leverage higher acceleration to reduce scan times, however, noise at these resolutions are already a concern. Fortunately, developments in patch-based methods have shown promise in reducing thermal noise in these multi-contrast images (Bazin et al, 2019). Here we evaluate a modification of the NORDIC fMRI denoising method (Dowdle et al, 2021, Vizioli, 2021), known as transformed domain NORDIC ("tNORDIC", Moeller et al 2023), on both 7T and 10.5T multi-echo GRE (ME-GRE) images with the goal of producing usable whole head quantitative, 0.37mm images with sub-10 minute acquisition times.

Following testing on publicly available data (Gulban et al, 2022), we acquired whole head, highly accelerated multi-echo GRE images at 7T (4 runs, TA:9:21, 0.37x0.37x0.37mm3, GRAPPA 3x3, 6 TEs: 4.23-24.98ms; TR:31ms, FA:11°, 7/8ths partial Fourier, 2xAP, 2xRL) and 10.5T (2 runs, TA:9:57, 0.37x0.37x0.37mm3, GRAPPA 3x3, 7 TEs:4.07-27.35ms; TR:33ms, FA:10°, 7/8ths partial Fourier, 1xAP, 1xRL). We evaluated several aspects of the data: 1) improved signal quality from reduced variance, 2) improvements in quantitative fits, 3) preservation of fine detail and 4) consistency in results compared to an average of multiple runs. We applied tNORDIC to the magnitude images and compared against alternatives: patched based method, LCPCA (Bazin et al, 2019), and non-local means (NLM) denoising (Avants et al, 2021). Alignment between the separate runs was calculated using antsRegistration. We estimated the quantitative T2* fit using a log-linear approach (t2smap, tedana (DuPre et al, 2021). T2* estimates and a T2*-based weighted average, were compared between the denoised data and 1 run or the average of 4 runs of the original data. T2* values were extracted from a gray matter (GM) ROI generated by manual segmentation.

After tNORDIC, the data from a single echo are less corrupted by thermal noise compared to the original images (7T data shown, Figure 1A). The image quality following tNORDIC matches that of the 4-run average (Figure 1A). In our hands, the alternative methods (LCPCA and NLM) produce blurring in the data, obscuring details (Figure 1A, right). The resulting estimates of T2* parameters are less noisy after tNORDIC (Figure 1B), reproducing the values observed after averaging 4 runs of the original data. A minimum intensity projection of the weighted average show that tNORDIC data preserves details (Figure 1C). Within the gray matter mask, one run of data after tNORDIC is more consistent run-to-run (i.e. test/re-test, Figure 2A) and more consistent with the 4-run average (Figure 2B). After tNORDIC, one run has a stronger correlation with the estimates derived from the 4-run average (r = 0.22, p <<0.001) compared to the original data (r = 0.12, p<<0.001). For the 10.5T data we found an additional benefit, in that tNORDIC recovers signal in the last echo which would be otherwise buried in noise due to shorter T2* values.

These results show that, after tNORDIC denoising, a single sub-10 minute acquisition can produce high quality, whole brain 0.37mm isotropic, quantitative data at 10.5 and 7T field strengths. The shortened scan times mitigate motion sensitivity and improve subject comfort. Here we have applied tNORDIC to ME-GRE magnitude images, but this method is applicable to any short series such as the ME-MP2RAGE sequence or distortion matched T1w-EPI. Further work will consider how this processing improves phase data for quantitative susceptibility mapping (QSM).

Anatomy and Functional Systems

Cortical Anatomy and Brain Mapping

Neuroanatomy Other ²

Anatomical MRI ¹

HIGH FIELD MR

MRI

STRUCTURAL MRI

Structures

Sub-Cortical

Workflows

Other - Denoising