Optimizing intersubject alignment of imaging data using anatomical and functional landmarks
Poster No:
1968
Submission Type:
Abstract Submission
Authors:
Jayson Jeganathan1, Bryan Paton2, Nikitas Koussis2, Michael Breakspear3
Institutions:
1The University of Newcastle, New Lambton Heights, NSW, 2University of Newcastle, New Lambton Heights, New South Wales, 3University of Newcastle, Newcastle, N/A
First Author:
Co-Author(s):
Introduction:
Traditional neuroimaging alignment methods map from an individual's brain to a template based on anatomical landmarks. If the anatomical locus supporting a particular function differs across individuals, functional imaging data will be misaligned by these methods. Functional alignment accounts for these differences by aligning imaging data based on functional landmarks, potentially increasing power in group MRI studies. However, alignment based on fMRI can discard useful anatomical constraints on alignment. This may be problematic in brain regions with poor fMRI signal-to-noise ratio. We tested whether an optimised combination of anatomical and functional alignment could balance the advantages of each approach and improve classification accuracy in a decoding framework.
Methods:
We analysed 3T fMRI from 100 Human Connectome Project participants consisting of 18 contrasts from 7 different tasks. The effectiveness of alignment methods was measured by classification accuracy within a decoding framework. A linear support vector classifier was trained to decode task labels from task fMRI maps, and tested with cross-validation. Standard anatomical alignment mapped from vertex loci in source subjects to loci in the target subject (or template brain) based on anatomical similarity1. For functional alignment, a linear transformation mapped anatomical vertices in source subjects to vertices in the target subject on the basis of similar fMRI response to 1 hour of movie-viewing. This mapping captured individually specific functional localisations (Figure 1). We considered several functional alignment methods including the Procrustes method and ridge regression. In this work, we extended the functional alignment method by mapping from a source subject to a regularized target consisting of a combination of the source subject's (weighted by γ, where γ ϵ [0,1]) and the original target's functional data (1-γ). When γ=0, the method corresponds to pure functional alignment. When γ=1, the method is an identity mapping, which is equivalent to anatomical alignment alone. Values of 0 < γ < 1 optimise both anatomical landmarks and functional co-activation.
Results:
Alignment between brain maps was optimized for intermediate parameter values 0.2-0.5, demonstrating the advantage of combining anatomical and functional features (Figure 2a). The improvement in task classification accuracy when interpolating between anatomical and functional methods was robust across a range of common functional alignment methods . The optimal parameter value, calculated from one cohort, generalized to other cohorts (Figure 2b). Results were similar when resting state functional connectivity rather than movie-viewing fMRI was used for functional alignment.
Conclusions:
By combining anatomical and functional information, we accounted for individual heterogeneity in functional topographies while incorporating anatomical constraints. The method improved the overlap between inter-individual brain maps beyond either anatomical or functional alignment alone, hence improving the predictive capacity of functional brain maps in a decoding framework. Modelling individual differences in this way may increase power in group MRI studies and reduce the sample sizes needed for clinically useful findings. Our findings demonstrate that macro-anatomy provides a partial lens into the inherent variability of individual neural landscapes.
Modeling and Analysis Methods:
Activation (eg. BOLD task-fMRI)
Classification and Predictive Modeling 2
Methods Development
Multivariate Approaches 1
Keywords:
Other - functional alignment
1|2Indicates the priority used for review