Lowering the Thermal Noise Barrier in Functional Brain Mapping with Magnetic Resonance Imaging

Luca Vizioli, Steen Moeller, Logan T. Dowdle, Mehmet Akçakaya, Federico De Martino, Essa Yacoub, Kâmil UǧurbilbioRxiv 2020 · DOI

TL;DR

This paper introduces NORDIC, a novel fMRI denoising technique that effectively suppresses thermal noise, a major limitation in high-resolution brain imaging. NORDIC significantly improves signal-to-noise ratio (SNR) and functional activation detection without sacrificing spatial precision, opening doors for more detailed and reliable brain mapping.

What Problem Does This Paper Address?

Functional Magnetic Resonance Imaging (fMRI) is a cornerstone for understanding brain activity, but its utility is often hampered by a poor signal-to-noise ratio (SNR). This is particularly true for high-resolution imaging (both spatial and temporal), where thermal noise becomes a dominant factor. This noise, which is indistinguishable from random Gaussian noise, obscures subtle neural signals, making it difficult to accurately detect and map brain activity. Existing methods often struggle to effectively remove this thermal noise without introducing unwanted spatial blurring, thus compromising the very precision that high-resolution imaging aims to achieve.

Methods

The core of this paper is the introduction of the NORDIC (NOise Reduction with Distribution Corrected) PCA technique. Unlike traditional methods that might focus on structured noise, NORDIC specifically targets and suppresses thermal noise, which is characterized as zero-mean, Gaussian-distributed noise.

The NORDIC algorithm operates on complex-valued fMRI data and employs a locally low-rank (LLR) model. The key steps involve:

  1. g-factor normalization: The data is scaled by the g-factor, which accounts for the spatially varying amplification of noise introduced by parallel imaging. This ensures zero-mean and spatially identical noise within local patches.
  2. Patch extraction and Casorati matrix formation: The fMRI time series is divided into small, overlapping patches, and these patches are vectorized to form a Casorati matrix.
  3. Singular Value Thresholding (SVT): A singular value decomposition is performed on the Casorati matrix. Singular values below a noise-level-determined threshold are set to zero, effectively removing noise components.
  4. Patch averaging: The denoised patches are averaged to reconstruct the denoised image series.
  5. g-factor de-normalization: The denoised data is multiplied back by the g-factor map to restore signal intensities.

A crucial aspect of NORDIC is its parameter-free threshold selection, which is determined empirically from the noise level, making it robust and less prone to user bias.

Datasets

The study utilized fMRI data from four healthy right-handed subjects. The datasets included:

  • 7 Tesla (7T) data:
    • 0.8 mm isotropic resolution with a block design visual stimulation paradigm.
  • 0.5 mm isotropic resolution with a block design visual stimulation paradigm.
  • 1.6 mm isotropic resolution with a fast event-related design (face detection task).
  • 3 Tesla (3T) data:
    • Standard HCP acquisition parameters (2 mm isotropic resolution, block design visual stimulation).
  • Modified HCP protocol (1.2 mm isotropic resolution, longer TR).
  • Resting-state fMRI data.

The paper also included comparisons with other denoising methods using 0.5 mm isotropic resolution data.

Key Insights and Ideas

  • Thermal noise is a significant barrier: The paper strongly emphasizes that thermal noise, often overlooked in favor of structured noise, is a major impediment to high-resolution fMRI, especially at higher field strengths.
  • NORDIC's targeted approach: The key innovation is NORDIC's ability to specifically target and suppress thermal noise without affecting the underlying signal or introducing spatial blurring. This is a significant departure from traditional spatial filtering methods.
  • Preserving spatial precision is paramount: The authors highlight that while SNR improvement is desirable, it should not come at the cost of spatial resolution. NORDIC achieves this balance, making it particularly valuable for fine-grained brain mapping.
  • Broad applicability: The study demonstrates NORDIC's effectiveness across various field strengths (3T and 7T), spatial resolutions (from sub-millimeter to supra-millimeter), cortical regions, stimulation paradigms (block design, event-related, resting state), and acquisition strategies.

Key Results

  • Significant tSNR improvement: NORDIC consistently improved temporal SNR (tSNR) by more than 2-fold compared to standard reconstruction across all subjects and runs.
  • Enhanced activation detection: Functional maps generated with NORDIC showed a significantly larger extent of activated voxels at the same t-threshold compared to standard methods. A single NORDIC run could achieve comparable or better results than concatenating 3-5 standard runs.
  • Preservation of spatial precision: NORDIC did not introduce spatial blurring, as evidenced by functional point spread function (PSF) measurements and the preservation of fine image features.
  • Comparable to or better than spatial smoothing: NORDIC achieved similar or better improvements in t-statistics compared to spatial smoothing, but without the associated loss of spatial resolution.
  • Robustness across conditions: NORDIC demonstrated consistent benefits across different field strengths, resolutions, and experimental paradigms, including resting-state fMRI.
  • Comparison with other methods: NORDIC outperformed global PCA (PCAwn) and DWIdenoise (MPPCA) in terms of image quality, t-statistics, and spatial smoothing, particularly for challenging high-resolution data.

Limitations and Future Work

  • Computational cost: While NORDIC offers significant advantages, the paper notes that it can be computationally intensive, especially for very long fMRI runs. However, they mention that patch averaging with overlap can mitigate this.
  • Hyperparameter selection for patch size: While the noise threshold is automatic, the patch size selection has some heuristic basis, and the paper acknowledges that fulfilling the "no two voxels within a patch are unaliased" constraint can be challenging for very long fMRI runs.
  • Further exploration of resting-state fMRI: The paper presents preliminary results for resting-state fMRI and suggests that a more in-depth analysis is required to fully appreciate NORDIC's impact on these types of acquisitions.
  • Potential for combined gains: The authors suggest that combining NORDIC with advancements in acquisition methods (e.g., higher magnetic fields, higher channel count RF coils) could lead to even greater improvements in spatial resolution and data quality.

Summary updated

Abstract

Abstract Functional magnetic resonance imaging (fMRI) has become one of the most powerful tools for investigating the human brain. However, virtually all fMRI studies have relatively poor signal-to-noise ratio (SNR). Here we introduce a novel fMRI denoising technique, which suppresses noise that is indistinguishable from zero-mean, Gaussian-distributed noise. Thermal noise, falling in this category, is a major source of noise in fMRI, particularly, but not exclusively, at high spatial and/or temporal resolutions. Using 7-Tesla high-resolution data, we demonstrate improvements in temporal-SNR, the detection of stimulus-induced signal changes, and functional maps, while leaving stimulus-induced signal change amplitudes, image spatial precision, and functional point-spread-function unaltered. We also show that the method is equally applicable when using supra-millimeter resolution 3- and 7-Tesla fMRI data, different cortical regions, stimulation/task paradigms, and acquisition strategies. This denoising approach improves key metrics of functional activation detection while preserving spatial precision.
BibTeX
@misc{vizioli2020lowering,
  author = {Luca Vizioli and Steen Moeller and Logan T. Dowdle and Mehmet Akçakaya and Federico De Martino and Essa Yacoub and Kâmil Uǧurbil},
  title = {{Lowering the Thermal Noise Barrier in Functional Brain Mapping with Magnetic Resonance Imaging}},
  year = {2020},
  doi = {10.1101/2020.11.04.368357},
  howpublished = {bioRxiv},
}
Figure