TL;DR
This paper introduces a new fMRI analysis technique, source-localized phase regressor (sPR), to mitigate the bias of standard BOLD fMRI towards large draining veins. By applying sPR to semantic category localization tasks (face and place recognition), they found that a significant portion of previously identified brain regions were actually driven by these large veins, leading to a re-evaluation of semantic category representation, particularly its hemispheric lateralization.
What Problem Does This Paper Address?
Gradient-echo BOLD fMRI, a workhorse for neuroimaging, has a known limitation: it's disproportionately sensitive to signals from large draining veins rather than the microvasculature directly associated with neural activity. This "large vein bias" can distort the localization and characterization of brain regions, especially when inferring neural differences from BOLD signal variations. Specifically, this paper investigates how this bias affects the identification and understanding of brain areas involved in semantic category recognition (like faces and places). Previous studies have suggested that these semantic areas are often right-hemisphere dominant and relatively rare. However, the authors hypothesize that this observed pattern might be an artifact of the BOLD signal's vascular bias, potentially overemphasizing regions with larger draining veins.
Methods
The core innovation is the source-localized phase regressor (sPR) technique. This method builds upon existing phase-regressor (PR) techniques but introduces two key improvements:
- Unbiased Least Squares Loss Function: Unlike previous methods that used a chi-squared loss function (which could lead to over-correction and suppression of non-vein signals), sPR employs an unbiased ordinary least squares (LOLS) loss function. This aims for more accurate estimation of the large vein contribution.
- Source Localization for Phase Information: sPR leverages the fact that large veins create distinct magnetic field distortions. While voxels directly within a large vein might have low phase signal-to-noise ratio (fSNR), adjacent voxels often have high fSNR. sPR intelligently uses this phase information from neighboring voxels to suppress the large vein signal, even in voxels with poor intrinsic phase information.
The authors validated sPR through:
- Computer Simulations: They simulated BOLD activity under various signal-to-noise conditions to demonstrate sPR's robust and unbiased suppression of large vein contributions compared to the original PR method.
- fMRI Data Analysis: They applied sPR to fMRI data from human subjects performing semantic category localization tasks (identifying faces and places). They then compared the results obtained using standard magnitude-based BOLD analysis with those from sPR-processed data.
Datasets
The study utilized fMRI data collected from four human subjects (all male, mean age 28) at UC Berkeley. The data was acquired using a 4 Tesla Varian INOVA scanner. For the semantic category localization task, stimuli were drawn from the public Caltech 256 database, comprising images from three categories: inanimate objects, human faces, and places.
Key Insights and Ideas
- Large Veins Significantly Inflate Semantic Area Size: A substantial portion of the voxels identified as belonging to face and place processing areas were found to be primarily driven by large draining veins. This means our understanding of the spatial extent of these areas has likely been overestimated.
- Hemispheric Lateralization May Be an Artifact: The observed right-hemisphere dominance for semantic processing, particularly in the Parahippocampal Place Area (PPA), appears to be at least partially explained by a bias towards larger veins in the right hemisphere. When this bias is removed, PPA shows a tendency towards left-hemisphere lateralization.
- ROI Size and fSNR Correlate with Vein Contribution: Regions that are larger and have higher fSNR (indicating stronger BOLD signals) tend to have a greater contribution from large draining veins. This suggests that the very properties that make a region detectable with standard fMRI might also make it more susceptible to vein bias.
Key Results
- Quantified Vein Contributions: The study estimated that:
-
38% of right PPA voxels
- 14% of left PPA voxels
- 16% of right FFA voxels
- 6% of left FFA voxels predominantly reflect signal from large draining veins.
-
Re-evaluation of PPA Lateralization: Without large vein suppression, PPA was right-hemisphere lateralized in both size and fSNR. After sPR, PPA became left-hemisphere lateralized in size and fSNR.
- FFA Lateralization Reduced: While FFA remained right-hemisphere lateralized after sPR, the degree of lateralization was reduced, suggesting a similar, though less pronounced, vein bias effect.
- Correlation Findings: Large vein contribution was positively correlated with both ROI size and fSNR.
Limitations and Future Work
- Neighborhood Size in sPR: The study used a seven-voxel neighborhood for sPR in the fMRI analysis. The authors acknowledge that future work is needed to determine the optimal neighborhood size and whether kernel-weighted neighborhoods would be more effective.
- Generalizability to Other Regions/Tasks: While the study focused on FFA and PPA, the authors suggest that future applications of sPR could reveal new semantic category areas previously obscured by large vein bias and improve the functional characterization of visual cortex more broadly.
- Higher Field Strengths: The authors note that while higher magnetic field strengths can reduce some vein bias, gradient-echo BOLD fMRI still suffers from it. They propose that sPR's ability to utilize phase information makes it particularly promising for future studies at higher field strengths and resolutions, potentially leading to even more efficient suppression of large vein bias.
- Understanding Venous Functional Organization: The paper highlights the poor understanding of cortical vasculature organization and its relationship to neurons. Future research into this "venous functional organization" could further refine our interpretation of fMRI signals.
Summary updated