Source-linked AI summary
Connectopic mapping with resting-state fMRI
Koen V. Haak, Andre F. Marquand, Christian F. Beckmann
TL;DR
Fine-grained patterns of connectivity and the spatial layout of functional anatomy remain markedly lacking. This paper proposes a principled, fully data-driven connectopic mapping approach with spatial statistical inference, demonstrating biologically meaningful, individualized, overlapping connectopies and significant identification rates.
Problem
Fine-grained connectivity patterns and the spatial layout of the brain’s functional anatomy remain markedly lacking, motivating improved connectopic mapping.
Method
The approach combines spectral embedding of voxel-wise connectivity fingerprints with trend-surface-based spatial statistical inference in resting-state fMRI.
Results
55-63% success rates were achieved against a 1.67% chance level, while the approach produced biologically meaningful, individualized, overlapping connectopies.
Takeaways & Limitations
The framework supports inference over connectopies and a multivariate characterization of spatial functional-connectivity topography.
Takeaways & Limitations
Existing approaches can erroneously uncover a single superposition when multiple connectopies simultaneously exist within the area of interest.
Abstract
from arXiv · showhide
Brain regions are often topographically connected: nearby locations within one brain area connect with nearby locations in another area. Mapping these connection topographies, or 'connectopies' in short, is crucial for understanding how information is processed in the brain. Here, we propose principled, fully data-driven methods for mapping connectopies using functional magnetic resonance imaging (fMRI) data acquired at rest by combining spectral embedding of voxel-wise connectivity 'fingerprints' with a novel approach to spatial statistical inference. We applied the approach in human primary motor and visual cortex, and show that it can trace biologically plausible, overlapping connectopies in individual subjects that follow these regions' somatotopic and retinotopic maps. As a generic mechanism to perform inference over connectopies, the new spatial statistics approach enables rigorous statistical testing of hypotheses regarding the fine-grained spatial profile of functional connectivity and whether that profile is different between subjects or between experimental conditions. The combined framework offers a fundamental alternative to existing approaches to investigating functional connectivity in the brain, from voxel- or seed-pair wise characterizations of functional association, towards a full, multivariate characterization of spatial topography.
1. Introduction
Brain connectivity is often organized as gradual spatial topographies, but existing methods inadequately capture fine-grained and overlapping connectopies. The paper proposes a data-driven manifold-learning framework with spatial-statistical inference to map and compare these patterns.
- Connectopies describe orderly connectivity in which nearby brain locations connect with nearby locations elsewhere in the brain.
- Fine-grained connectivity patterns are difficult to characterize, while parcel-based approaches often assume homogeneous connectivity.
- Overlapping connectopies can coexist within one brain area, creating a major obstacle for accurately characterizing connectivity organization.
- Moving-seed methods may recover a single superposition instead of the true multiplicity of coexisting organizational modes.
- The proposed approach computes similarities among voxel-wise functional-connectivity fingerprints and applies manifold learning to identify overlapping connectopies.
- Trend surface analysis reduces connectopic patterns to a small parameter set, enabling statistical inference and comparisons across subjects or experimental conditions.
- Applied to resting-state fMRI, the framework produced biologically valid overlapping connectopies in a fully data-driven manner.
2. Methods
The framework maps connectopies from resting-state fMRI using connectivity fingerprints, manifold learning, and spatial statistics. It represents similar connectivity profiles as spatially organized embeddings and supports inference over the resulting connectopic maps.
- Framework: The framework combines connectivity fingerprinting, manifold learning, and spatial statistics to map connectopies from resting-state fMRI.It was demonstrated on somatotopic organization in M1 and retinotopic organization in V1 using data from 60 HCP subjects.
- Connectivity fingerprinting: Connectivity fingerprints correlate each ROI voxel’s time series with SVD-transformed time series from gray-matter voxels outside the ROI.The resulting matrix C contains one correlation map for each ROI voxel.
- Manifold learning: The η2 coefficient measures similarity between voxel-wise connectivity profiles, ranging from 0 for entirely dissimilar to 1 for entirely similar profiles.The similarity matrix is transformed into a connected graph before manifold learning.
- Manifold learning: A single connected graph is enforced so the ensuing connectopies cover the entire ROI rather than separate restricted components.The connectivity graph threshold ε is set to the minimum value required for graph connectivity.
- Spatial Statistics: The spatial-statistics model approximates each connectopic map with a polynomial spatial trend plus a Gaussian process for finer spatial variation.The resulting parsimonious coefficients can be tested parametrically or non-parametrically or used as features in other analyses.
3. Results
Resting-state connectopic mapping recovered biologically plausible topographies in motor and visual cortex, including overlapping modes of organisation. Individual M1 maps were reproducible across sessions, subject-specific, and robust to modest ROI inaccuracies.
- Connectopic mapping at the group-level: The dominant group-level connectopy in M1 corresponded clearly to its somatotopic map.It showed mirror-symmetric interhemispheric connectivity and topographically organised connectivity with anterior cerebellum.
- Connectopic mapping at the group-level: In V1, the dominant and second-dominant connectopies followed distinct eccentricity and polar-angle trajectories.This demonstrates that multiple overlapping topographic organisations can coexist within one area.
- Connectopic mapping at the group-level: Superimposing V1’s two dominant connectopies produced a reproducible but biologically invalid diagonal gradient that could yield nonsensical parcellations.Methods that account for overlapping connectopies avoid treating this superposition as a single valid organisation.
- Connectopic mapping in single subjects: Individual M1 connectopies resembled group maps and were highly similar across sessions, subjects, and modest ROI-definition changes.The maps were also supported by high cross-session stability in scans as short as approximately 7.5 minutes with a TR of approximately 2 seconds.
- Connectopic mapping in single subjects: Local nonlinear manifold learning produced biologically plausible connectopies, whereas linear methods produced disorganised maps and Isomap did not disentangle overlapping connectopies.The results support preferring locally nonlinear approaches such as Laplacian eigenmaps for these data.
- Connectopic mapping in single subjects: M1 connectopies showed greater within-subject than between-subject reproducibility, indicating subject-specific functional-connectivity organisation.Spatial models correctly identified 62% of left-ROI and 53% of right-ROI runs, versus a 1.67% chance level.
4. Discussion
The paper presents a fully data-driven resting-state fMRI framework that maps biologically meaningful, individualized, and overlapping connectopies while enabling spatial statistical inference. Its demonstrations support nonlinear manifold learning, compact spatial modeling, and applications beyond stimulus- or task-based analyses.
- Connectopic mapping: Resting-state fMRI connectopic mapping produced biologically meaningful, individualized connectopies, including multiple overlapping topographic organizations within one area.The approach addresses overlapping connectopies that earlier methods could not simultaneously identify.
- Spatial statistical inference: Trend-surface analysis condensed high-dimensional connectopy images into few coefficients, enabling tests of spatial connectivity variation across subjects and conditions.This parameterization also supports anatomically relevant hypotheses beyond traditional voxel- or cluster-wise testing.
- Spatial statistical inference: The spatial statistical model compactly described connectopic maps, achieving 55-63% mate-based retrieval success against a 1.67% chance level.The experiment was confined to the human motor strip, yet the identification rates were reported as high.
- Connectopic mapping: Nonlinear manifold learning was preferred because a linear approach produced disorganized, biologically implausible connectopies.The demonstrated implementation used local nonlinear manifold learning to capture fine-grained topographic structure.
- Biological validation: Estimated motor-cortex connectopies matched topographic connections with the opposing hemisphere and anterior cerebellum, supporting their biological plausibility.Projecting a map from one brain region across the brain was proposed as a generic way to discover new topographically connected information-processing networks.
- Applications: The framework can investigate association-cortex topographic maps and may support translational studies when stimulus-driven or task-based experiments are precluded.It does not require prior knowledge of the investigated area's topographic organization and may provide more sensitive disease markers.