(D-140) Noise Reduction of Visual Task-Based Functional Magnetic Resonance Imaging using Noise Reduction with Distribution Corrected Principal Component Analysis

Introduction:: Low signal-to-noise ratio (SNR) prevents high resolution fMRI from precisely locating brain structures, hindering the utility prospects for clinical and neuroscientific applications. We aimed at pushing the spatial resolution to sub-millimeter acquisitions and also implemented a denoising method, Noise Reduction with Distribution Corrected (NORDIC) PCA, to suppress noise of the image series that is indistinguishable from thermal noise. This work highlights the ability to detect brain activity and the performance of Noise Reduction with Distribution Corrected to denoise by comparing quantitative evaluations to the standard manufacturer reconstruction for visual task-based high-resolution fMRI datasets.

Materials and Methods::

Ten volunteers (mean age 24/2.977 std, 3 female 7 male) participated in a visual task involving a visual checkerboard block design displayed on an MR compatible screen. The task was run eight times with four tasks being run to collect 1mm isotropic resolution images and the other four runs being run to produce 0.8mm isotropic resolution images. The task includes four blocks that started with a flickering checkerboard (8Hz) for 20s and a 20s white fixation following the initial sequence. Each subject repeated the visual task fMRI experiment on a 7T Siemens Magneton Terra to collect functional images.

MATLAB (R2019b; the Mathworks, Inc, Natick, MA, USA) and the software Statistical Parametric Mapping (SPM) was utilized to preprocess and analyze the functional MRI data across the eight runs for the two reconstruction methods. Reconstructed fMRI timeseries were motion realigned and no spatial smoothing was performed. A General Linear Model with visual stimulation predictors analyzed the preprocessed images separately for each run and reconstruction technique. A one-sample t-statistic across all eight runs for each reconstruction modality revealed significant voxels and consistency, and the t-statistic map produced two regions of interests for the left and right visual cortex for the NORDIC and standard reconstruction data to compute the temporal SNR (tSNR). To evaluate tSNR differences between different spatial resolution data as well as NORDIC and Standard reconstruction methods, one-way repeated measures analyses of variance (ANOVAs) were conducted.

Results, Conclusions, and Discussions::

The ANOVA analysis revealed differences between 1mm isotropic resolution images tSNR and 0.8 mm isotropic resolution images tSNR. Across the two reconstruction methods, the post hoc analysis with Bonferroni test features better tSNR in the 1mm isotropic resolution images compared to the 0.8mm isotropic resolution images (for NORDIC, mean difference of tSNR=3.5, p < 0.4. For Standard, mean difference of tSNR=13.206, p < 0.001; Table 1). As for the comparison between the NORDIC and Standard, the ANOVA analysis revealed a significant difference between the two reconstruction methods. The post hoc analysis with Bonferroni test shows that the NORDIC has significantly higher tSNR compared to the tSNR in the Standard (for 1mm isotropic resolution images, mean difference of tSNR=17.119, p < 0.001. For 0.8mm isotropic resolution images, mean difference of tSNR=26.825, p < 0.001; Table 1).

Finding lower tSNR in 0.8mm compared to 1mm was expected, and the results comparing NORDIC and Standard reconstruction were also expected. We hypothesized NORDIC to have significantly higher tSNR than the standard (for 1mm isotropic resolution images, the total mean of Standard tSNR=43.683 and the total mean of NORDIC=60.803. For 0.8mm isotropic resolution images, the total mean of Standard=30.787 and the total mean of NORDIC=57.302; Table 2). In future studies, we plan to correct and optimize aspects of the code to achieve precise and accurate results.

We demonstrated the 7T sub-millimeter detection of brain activity is feasible with sufficient tSNR and statistical significance. Based on our findings, we hope our techniques and methods can be utilized in neuroscientific and clinical applications. Our approach to utilizing reconstruction techniques may offer new perspectives or opportunities to push the boundaries of collecting significantly more precise images of brain function. Compared to standard methods, we hope to push the resolution used in current fMRI to find activated regions within separate layers of gray matter of the brain. .

Acknowledgements (Optional): :

References (Optional): :