TDM: a temporal decomposition method for removing venous effects from task-based fMRI

Kendrick Kay, Keith Jamison, Ru‐Yuan Zhang, Kâmil UǧurbilbioRxiv 2019 · DOI

TL;DR

This paper introduces Temporal Decomposition through Manifold Fitting (TDM), a novel analysis method for fMRI that disentangles brain activity signals originating from small blood vessels (microvasculature) and large veins (macrovasculature). By doing so, TDM significantly improves the spatial accuracy of fMRI, reduces artifacts, and offers a more precise understanding of neural activity, especially in high-resolution imaging.

What Problem Does This Paper Address?

Functional Magnetic Resonance Imaging (fMRI) is a powerful tool for studying brain activity, but its reliance on the Blood Oxygenation Level Dependent (BOLD) signal, particularly with gradient-echo sequences, is hampered by venous effects. These effects, caused by blood flowing through veins, can distort the BOLD signal, leading to inaccurate localization of neural activity (up to 4mm displacement) and limiting spatial resolution. While alternative methods like spin-echo sequences can mitigate these venous effects, they often come with reduced signal-to-noise ratio and coverage. Existing analytical workarounds, such as masking veins or focusing on deeper cortical layers, sacrifice valuable information. This paper tackles the challenge of obtaining spatially accurate and venous-free BOLD signals from standard fMRI acquisitions.

Methods

The core of this paper is the introduction of the Temporal Decomposition through Manifold Fitting (TDM) method. TDM is a data-driven approach that works by:

  1. Principal Component Analysis (PCA): Analyzing the timecourses of BOLD responses across voxels to identify the main axes of variation. This reduces the dimensionality of the data.
  2. Manifold Fitting: Projecting these timecourses into a 3D space and identifying a 1-dimensional manifold (essentially a line or arc) that captures the primary variation. This manifold is interpreted as representing a spectrum from early (microvasculature) to late (macrovasculature) hemodynamic responses.
  3. Decomposition: Using the identified "Early" and "Late" timecourses as basis functions in a General Linear Model (GLM) to decompose the observed BOLD response in each voxel into components attributed to the microvasculature and macrovasculature, respectively.

This method is designed to be compatible with standard gradient-echo fMRI and can be applied retrospectively to existing datasets.

Datasets

The study utilized data from four experiments (E1-E4) involving eleven subjects. These experiments employed different fMRI acquisition protocols:

  • Experiment E1: High-resolution (7T, 0.8 mm) gradient-echo for eccentricity stimuli.
  • Experiment E2: High-resolution (7T, 0.8 mm) gradient-echo for category stimuli.
  • Experiment E3: Spin-echo (7T, 1.05 mm) for eccentricity stimuli.
  • Experiment E4: Low-resolution (3T, 2.4 mm) gradient-echo for eccentricity stimuli.

A total of sixteen datasets were collected across these experiments. The paper also references data from previous publications for comparison.

Key Insights and Ideas

  • Two distinct hemodynamic timecourses: The TDM method consistently identifies two primary timecourses that explain the majority of variation in fMRI responses: an "Early" timecourse associated with microvasculature (capillaries and small venules) and a "Late" timecourse linked to macrovasculature (veins).
  • Veins cause delayed and spatially dispersed signals: The "Late" timecourse, attributed to veins, is characterized by larger amplitudes, delayed peaks, and a tendency to be more prominent in outer cortical depths and areas with lower EPI intensity. This directly explains the spatial inaccuracies observed in standard fMRI.
  • TDM improves spatial accuracy: By separating these components, TDM effectively removes the venous-induced artifacts, leading to more accurate localization of neural activity and flatter depth-dependent response profiles.
  • Data-driven flexibility: TDM's data-driven nature means it adapts to the specific characteristics of each dataset, making it robust across different acquisition parameters, subjects, and experimental designs.
  • Compatibility with various protocols: The method demonstrates effectiveness not only with high-resolution gradient-echo fMRI but also with lower-resolution and even spin-echo acquisitions, highlighting its broad applicability.

Key Results

  • TDM successfully decomposes BOLD responses into early (microvascular) and late (macrovascular) components.
  • The "Late" component consistently shows characteristics associated with venous blood flow: higher amplitude, delayed response, and bias towards outer cortical depths.
  • Applying TDM significantly reduces the superficial cortical depth bias present in gradient-echo BOLD responses.
  • TDM eliminates artifacts in cortical activity maps that are not present with standard GLM analysis.
  • The method is robust and yields consistent results across multiple datasets and subjects, even at resolutions as low as 2.4 mm.
  • TDM can be applied to spin-echo data to remove residual venous effects, suggesting it's a valuable tool even when specialized acquisition methods are used.

Limitations and Future Work

  • Sensitivity Loss: TDM involves decomposing responses into two often correlated timecourses, which can lead to a penalty in terms of higher variance in beta estimates compared to simpler models. For applications where sensitivity is paramount and specificity is less critical, TDM might not be the optimal choice.
  • Intrinsic Physiological Limitations: The effectiveness of TDM is inherently limited by the degree to which the microvascular and macrovascular timecourses are separable. In cases where these timecourses are very similar (e.g., in some subjects), the reduction in reliability can be more pronounced.
  • Vascular Complexity: TDM uses a two-component model for vascular contributions. The actual vascular system is more complex, and future work could explore more nuanced characterizations of vascular types and their dynamics.
  • Future Directions: The authors suggest further work on optimizing TDM for different experimental designs (e.g., condition duration), spatial/temporal resolutions, and amounts of data. Testing TDM on other types of experiments (sensory, cognitive, motor) and in other brain areas is also proposed. Additionally, comparing TDM with optimized VASO (Vascular Space Occupancy) sequences is suggested as a promising avenue for future research.

Summary updated

Abstract

Abstract Most functional magnetic resonance imaging (fMRI) is conducted with gradient-echo pulse sequences. Although this yields high sensitivity to blood oxygenation level dependent (BOLD) signals, gradient-echo acquisitions are heavily influenced by venous effects which limit the ultimate spatial resolution and spatial accuracy of fMRI. While alternative acquisition methods such as spin-echo can be used to mitigate venous effects, these methods lead to serious reductions in signal-to-noise ratio and spatial coverage, and are difficult to implement without leakage of undesirable non-spin-echo effects into the data. Moreover, analysis heuristics such as masking veins or sampling inner cortical depths using high-resolution fMRI may be helpful, but sacrifice information from many parts of the brain. Here, we describe a new analysis method that is compatible with conventional gradient-echo acquisition and provides venous-free response estimates throughout the entire imaged volume. The method involves fitting a low-dimensional manifold characterizing variation in response timecourses observed in a given dataset, and then using identified early and late timecourses as basis functions for decomposing responses into components related to the microvasculature (capillaries and small venules) and the macrovasculature (veins), respectively. We show that this Temporal Decomposition through Manifold Fitting (TDM) method is robust, consistently deriving meaningful timecourses in individual fMRI scan sessions. Moreover, we show that by removing late components, TDM substantially reduces the superficial cortical depth bias present in gradient-echo BOLD responses and eliminates artifacts in cortical activity maps. TDM is general: it can be applied to any task-based fMRI experiment, can be used with standard- or high-resolution fMRI acquisitions, and can even be used to remove residual venous effects from specialized acquisition methods like spin-echo. We suggest that TDM is a powerful method that improves the spatial accuracy of fMRI and provides insight into the origins of the BOLD signal.
BibTeX
@misc{kay2019tdm,
  author = {Kendrick Kay and Keith Jamison and Ru‐Yuan Zhang and Kâmil Uǧurbil},
  title = {{TDM: a temporal decomposition method for removing venous effects from task-based fMRI}},
  year = {2019},
  doi = {10.1101/868455},
  howpublished = {bioRxiv},
}
Figure