Hydra Nordic: A thermal-noise removal strategy for multi-echo fMRI

1698 

Abstract Submission 

Marco Flores-Coronado¹, Cesar Caballero-Gaudes²

¹Basque Center on Cognition, Brain and Language, Donostia-San Sebastián, Guipuzkoa, ²Basque Center of Cognition, Brain and Language, San Sebastián, Spain

Marco Flores-Coronado  
Basque Center on Cognition, Brain and Language
Donostia-San Sebastián, Guipuzkoa

Cesar Caballero-Gaudes  
Basque Center of Cognition, Brain and Language
San Sebastián, Spain

During MRI acquisition, artifacts (e.g., body movement, and respiratory effects)(Satterthwaite et al., 2019) and noise sources (e.g., thermal noise from the scanners)(Wald & Polimeni, 2017) hinder signal fidelity. Multi-echo fMRI techniques (ME), in turn, have demonstrated superior data quality compared to conventional single-echo acquisitions. ME combines signals from different echo times (TEs) to amplify the contrast-to-noise ratio(Poser et al., 2006). Complementing this, multi-echo independent component analysis (ME-ICA) refines signal quality by removing components that do not follow BOLD assumptions(Community et al., 2023; DuPre et al., 2021; Kundu et al., 2017). Additionally, Nordic has emerged as an effective method in reducing thermal noise during single-echo acquisition(Vizioli et al., 2021). Yet, an integrated approach leveraging the advantages of both ME and Nordic remains unformulated. Here, we investigate the optimal use of thermal-noise reduction using Nordic and ME-based techniques (e.g., optimal combination and ME-ICA). Our proposed method, named Hydra Nordic, explores the simultaneous use of these techniques for an enhancement in signal quality.

Five healthy volunteers (3 women) were scanned in a 3T Siemens PrimaFit MAGNETOM MR scanner using a 64-channel head at the Basque Center on Cognition, Brain and Language. Multi-echo fMRI data was collected during two resting state runs with different voxel resolutions (2.4 mm isotropic voxels: TR=1.2s, TE= 11.2/28.1/45/61.9 ms, and 2 mm isotropic voxels: TR=1.7s, TE=13.4/36.1/58.8/81.5 ms, both runs with SMS=5, GRAPPA=2, 65/75 whole-head sagittal slices respectively). Single-band reference (SBRef) images were acquired for each TE. T1-weighted MP2RAGE and T2-weighted Turbo Spin Echo images were collected at 1 mm voxel resolution in each subject. ME-fMRI data was minimally preprocessed with AFNI, including the removal of the 10 first volumes to achieve a steady-state magnetization, image realignment of the first echo to the SBRef and applying this transformation to the other echoes, and T2*-weighed linear (optimal) combination of the echoes using TEDANA (ME-OC). Before this standard ME-fMRI preprocessing, two versions of the Nordic denoising approach were also applied to reduce thermal noise: applying Nordic to each echo dataset independently (NORDIC-OC), and applying Nordic on the spatial concatenated dataset (3dZcat) from all the echoes and then separating them (3dZcutup) for subsequent preprocessing (HYDRA-OC). Both Nordic approaches used the noise-only scans (i.e., no RF excitation) acquired at the end of each run to estimate the hyper-parameters of the Nordic algorithm. Temporal signal-to-noise ratio (TSNR) maps (i.e. voxelwise mean divided by its standard deviation) were computed on the ME-OC, NORDIC-OC and HYDRA-OC datasets. We hypothesize that reducing thermal noise from echoes simultaneously would better capture thermal noise because we assume it to be similar between echoes

Temporal Signal-to-Noise Ratio (T-SNR) maps were computed from OC maps -TSNR=mean(OC)/sd(OC)- As previously reported, we found that Nordic reduced overall signal variance. Moreover, we found an positive increase in overall tSNR values between HYDRA-OC TSRN values as compared both to ME-OC, and NORDIC-OC. In FIgure 1, scatterplots from a representative subject show a positive increment in overall TSNR values (2.4mm isotropic voxel). In FIgure 2, scatter plots from the same individual inform that with lower voxel size, there is a greater increment in TSNR values between methods.

We argue that, concatenating the echo-time-series volumes let us better approach the thermal noise distribution. Furthermore, me found evidence that such estimation improves if voxel size decrease.

Exploratory Modeling and Artifact Removal ¹

Methods Development ²

FUNCTIONAL MRI

Workflows

Other - signal cleaning