Spurious correlations in surface-based functional brain imaging
Poster No:
1588
Submission Type:
Abstract Submission
Authors:
Jayson Jeganathan1, Nikitas Koussis2, Bryan Paton3, Richa Phogat3, James Pang4, Sina Mansour5, Andrew Zalesky6, Michael Breakspear3
Institutions:
1The University of Newcastle, New Lambton Heights, AK, 2University of Newcastle, New Lambton Heights, NSW, 3The University of Newcastle, New Lambton Heights, NSW, 4Monash University, Clayton, Victoria, 5National University of Singapore, Singapore, Singapore, 6The University of Melbourne and Melbourne Health, Melbourne, VIC
First Author:
Co-Author(s):
Introduction:
The study of functional MRI (fMRI) data is increasingly performed after mapping from volumetric voxels to surface vertices. Processing pipelines commonly used to achieve this mapping produce meshes with uneven vertex spacing. We investigated the causes and impacts of uneven vertex spacing.
Methods:
We used minimally pre-processed resting state fMRI data from the first 20 participants of the Human Connectome Project. Data were represented on the fsLR 32k surface. Statistical tests were corrected for spatial smoothness using the spin test.
Results:
Inter-vertex distances are correlated with sulcal depth, with up to 3 times greater inter-vertex distance in gyri than in sulci (r=0.526, p<0.001) (Figure 1). The bias is propagated from anatomical to empirical fMRI data, where the correlation between adjacent sulcal vertices becomes artefactually inflated due to their proximity (r=-0.508, p<0.001) (Figure 2). Biased nearest-neighbour correlations are present in all participants (one-sample t-test, t(19)=-41.183, p<0.001), on fsLR 32k and fsaverage meshes, with resting and movie-viewing MRI, and with MSMSulc and MSMAll. Biased neighbour correlations do not occur in volume data, demonstrating that the bias arises from surface processing alone (r=0.043, p=0.297). To reproduce the bias in-silico, we generated random Gaussian noise time series independently at each vertex of a participant's uneven surface mesh. Surface smoothing (2mm FWHM) alone yielded artefactually high neighbour correlations in sulcal vertices (r=0.812, p<0.001). We used the in-silico model to explore the consequences of biased neighbour correlations.
First, we found that functional parcellations are biased towards finding parcel boundaries at gyri, even when the underlying data is noise (t(29694)=41.636, p<0.001). Biased parcel boundaries are also seen in commonly used group parcellations (Schaefer et al., 2018) derived from empirical fMRI data (p<0.001). Second, we examined vertex-level fingerprinting accuracy, the ability to match an individual's functional connectome with their own re-test functional connectome. Even when test and re-test data are generated from independent uncorrelated Gaussian noise, individual-specific anatomical cortical folding information leaks into test and re-test functional connectomes, resulting in 98% mean fingerprinting accuracy (chance level 1/20 = 5%).
The onavg template was recently developed to reduce variability in inter-vertex spacing. When the template is projected to individuals' brain surfaces, we find a modest 20.7% reduction in variability. The onavg template abolishes gyral biases in functional parcellation boundaries, but worsens artefactual fMRI fingerprinting.
First, we found that functional parcellations are biased towards finding parcel boundaries at gyri, even when the underlying data is noise (t(29694)=41.636, p<0.001). Biased parcel boundaries are also seen in commonly used group parcellations (Schaefer et al., 2018) derived from empirical fMRI data (p<0.001). Second, we examined vertex-level fingerprinting accuracy, the ability to match an individual's functional connectome with their own re-test functional connectome. Even when test and re-test data are generated from independent uncorrelated Gaussian noise, individual-specific anatomical cortical folding information leaks into test and re-test functional connectomes, resulting in 98% mean fingerprinting accuracy (chance level 1/20 = 5%).
The onavg template was recently developed to reduce variability in inter-vertex spacing. When the template is projected to individuals' brain surfaces, we find a modest 20.7% reduction in variability. The onavg template abolishes gyral biases in functional parcellation boundaries, but worsens artefactual fMRI fingerprinting.
Conclusions:
Surface processing leads to a gyral bias in inter-vertex distances, resulting in gyrally biased fMRI correlations which can adversely impact subsequent analyses. These biases can be partially addressed using the onavg template. The impact of residual biases on one's pipeline can be quantified using surrogate noise data. Surface-based pipelines must be used with caution.
Modeling and Analysis Methods:
Methods Development 2
Motion Correction and Preprocessing 1
Segmentation and Parcellation
Task-Independent and Resting-State Analysis
Keywords:
Statistical Methods
Other - surface; parcellation;
1|2Indicates the priority used for review