Effects of NORDIC denoising on population receptive field maps

1327 

Abstract Submission 

David Linhardt¹, Garikoitz Lerma-Usabiaga², Michael Woletz¹, Christian Windischberger¹

¹High Field MR Center, Medical University of Vienna, Austria, ²Basque Center On Cognition, Brain and Language, BCBL, San Sebastian, Spain

David Linhardt  
High Field MR Center, Medical University of Vienna
Austria

Garikoitz Lerma-Usabiaga  
Basque Center On Cognition, Brain and Language, BCBL
San Sebastian, Spain

Michael Woletz  
High Field MR Center, Medical University of Vienna
Austria

Christian Windischberger  
High Field MR Center, Medical University of Vienna
Austria

In fMRI studies it is a common goal to achieve higher spatial resolution. However, measuring smaller voxels reduces the signal-to-noise ratio (SNR). NORDIC [1, 2] is a noise reduction method aiming to tackle this issue. Based on a patchwise principal component analysis (PCA), NORDIC identifies and removes components that are similar to thermal noise. This SNR increase goes hand in hand with an increase in image smoothness, however it was found that the increase is less than with other denoising methods [3]. In the special fMRI application of population receptive field (pRF) mapping, Gaussian smoothing of the image is directly corresponding to an increase in pRF size parameter estimation [4]. Within this work, we investigated the effects of NORDIC denoising on the resulting pRF parameter estimations.

We acquired fMRI data in three participants (two female, age: 20.3∓0.9) on a SIEMENS PrismaFit 3T scanner using the lower part of a 64-channel head coil. Subjects were scanned for one 5-minute functional run (TE/TR=1000/38ms, 1.5mm isotropic) while being presented a bar aperture (width=1.2°) moving through the visual field in eight different directions, revealing reversing checkerboards. The stimulated field of view was 9° radius. NORDIC denoising was applied on the NIFTI files after standard scanner reconstruction (NIFIT_NORDIC, github.com/SteenMoeller/NORDIC_Raw). Both the original and noise-reduced versions were minimally preprocessed using fMRIPrep v23.0.1 (fmriprep.org). The pRF mapping analysis was conducted on both conditions using the containerized solution prfprepare v1.3.5 (github.com/dlinhardt/prfprepare) and prfanalyze-vista v2.2.2_3.1.1 (github.com/vistalab/PRFmodel). Only voxels in V1-3 (Benson atlas [5]) and above the 20% variance explained threshold were used for the comparison. Assessment of NORDIC effects were performed in a voxel-by-voxel comparison.

The pRF position eccentricity and polar angle estimations did not show any systematic differences between the two conditions (Fig 1A and B), but including NORDIC in the preprocessing pipeline substantially increased the pRF size (Fig 1C). As expected for a noise-reduction method, NORDIC-processed results show higher variance-explained values (Fig 1D). To quantify the differences, we calculated Cohen's d effect sizes [6] for the different comparisons. These effect sizes confirm the quantitative results and yield no effects for the pRF center position (eccentricity 0.06; polar angle 0.01), a small effect in the pRF size (0.33) and a large effect for the variance explained (1.66).
Further quantification was conducted to assess the magnitude of pRF size increase. For that, we computed the ratio of pRF sizes estimation with and without NORDIC correction. As illustrated in Figure 2, a histogram of this ratio reveals a predilection for values greater than 1, substantiating an amplification in pRF size. The median of this distribution is positioned at 1.22, signifying a 22% increase in pRF size attributable to the removal of noise using NORDIC.

The application of NORDIC denoising in pRF mapping, results in a median 22% increase in pRF size estimations. These increases in pRF size estimation can not directly be linked to overall image smoothness (AFNI 3dFWHMx). Instead, we argue it arises from the integration of information from adjacent voxels with differing pRF characteristics, leading to peak broadening. This phenomenon is based on the inherent retinotopic organization of neighboring voxels. Previous studies have demonstrated that standard Gaussian spatial smoothing impacts pRF size [4], showing the described effect. Even though image smoothness is hardly increased using NORDIC [3], the effect of neighboring voxel integration most probably leads to the measured pRF size increases. Therefore, the use of NORDIC in preprocessing, despite its benefits in noise reduction, requires cautious interpretation due to its influence on pRF size estimations.

Activation (eg. BOLD task-fMRI) ¹

Other Methods

BOLD fMRI

Perception: Visual ²

FUNCTIONAL MRI

MRI

Vision

Other - pRF mapping, retinotopy, NORDIC