Functional connectivity of the insula in the resting brain

Study at a glance

• Authors: Franco Cauda, Federico D’Agata, Katiuscia Sacco, Sergio Duca, Giuliano Geminiani, Alessandro Vercelli
• Year: 2011
• Journal: NeuroImage
• DOI: 10.1016/j.neuroimage.2010.11.049
• N: 16
• Design: Resting-state fMRI (1.5T, 200 volumes); seed-based rsFC from ten insular ROIs plus connectivity-based parcellation (k-means, hierarchical, and voxelwise fuzzy c-means clustering)

Key findings

• Two complementary, anticorrelated insular networks. A ventral-anterior insula network couples to rostral anterior cingulate, middle/inferior frontal cortex and temporoparietal junction (a limbic/salience pattern); a dorsal-posterior insula network couples to premotor, sensorimotor and supplementary-motor cortex and mid-posterior cingulate (a sensorimotor-integration pattern).
• The anterior/posterior bipartition is confirmed by negative-correlation analysis: the two networks are mutually anticorrelated, and in a sample subject the anterior- and posterior-insula BOLD time courses correlate r = -0.47.
• Three independent parcellation methods (k-means, hierarchical, fuzzy c-means) converge on an anterior / posterior / transitional tripartition of the insula; the middle ROIs (6, 9) form the transitional cluster, and voxelwise fuzzy clustering reproduces the same ventral-anterior/dorsal-posterior split with an intermediate zone.
• Both networks are partially lateralized: the salience/anterior pattern connects more strongly to anterior cingulate on the right and to prefrontal cortex on the left; the posterior pattern connects more strongly to superior temporal and occipital cortex on the right. Right anterior ROIs are preferentially connected to brainstem, pons, right thalamus, right dlPFC and rACC — read as support for a right-insula role in arousal/attention.
• Subcortical connectivity follows a rostroventral→caudodorsal decorrelation gradient: positive coupling with basal ganglia and thalamus is strongest for the anterior maps (ROIs 1-6, 8) and falls toward the caudal/dorsal insula (ROIs 7, 9, 10).

Limitations

• 1.5T scanner, 4.5 mm slices; 5x5x5 mm seed ROIs exceed average cortical thickness (~3 mm), so signal may include adjacent (opercular, subcortical) structures — the authors test alternative ROI sizes/positions and report similar maps, but the concern is not eliminated.
• Resting-state correlations are non-directional and cannot say whether the anterior insula drives the posterior or vice versa, nor what the ‘salience’ vs ‘sensorimotor’ labels compute.
• Age-heterogeneous sample (mean ~53 y, SD 19; range spans young to elderly) — the authors show age is uncorrelated with head motion and use RFX/ANCOVA to guard against age effects, but the sample is not a young-adult norm.
• The anticorrelated (negative) networks depend on global-signal regression, whose validity for revealing true anticorrelation was contested at the time (Fox 2009; Weissenbacher 2009; Murphy 2009) — the authors flag this and treat the anticorrelation with caution.

Connections

• Methods: Resting-state functional connectivity
• Concepts: Insular cortex, Salience network, Default mode network, Anterior cingulate cortex (ACC), Von Economo neurons (VENs), Central autonomic network
• Debates: inferential?, Is interoception domain-general or organ-specific?
• Related studies: Craig (2002), Craig (2009), Farb et al. (2012), Haruki & Ogawa (2023), Brooks et al. (2005), Chen et al. (2021), Nord et al. (2021), Chang et al. (2013), Deen et al. (2011), Vercelli et al. (2015), Gong et al. (2015)

The wiki’s anterior/posterior insula dichotomy arrives on almost every page as anatomy — Craig’s lamina I afferent terminating posteriorly, re-represented anteriorly; the granular/agranular seam of the EPIC model; the cytoarchitectural tripartition (agranular anterior, dysgranular mid, granular posterior). Cauda et al. derive nearly the same split from a different kind of evidence entirely: the correlation structure of spontaneous BOLD fluctuations at rest, with no task, no stimulus, and no reference to cytoarchitecture. Ten seed ROIs drawn across the insular surface, 16 healthy adults lying still, and three clustering methods all return the same answer — the insula is functionally two things (with a transitional third), organized along its ventral-anterior → dorsal-posterior axis.

This is one of the founding resting-state demonstrations of that organization (with Taylor et al. 2009, Deen et al. 2011, and the group’s own companion papers), and it is why the dichotomy this wiki leans on is not merely an inference from Nissl stains.

The wiki holds the tripartite successor: Chang et al. (2013) parcellate the right insula the same way (resting-state connectivity, k-means) but let a validity index choose the cluster number rather than fixing it, and land on k=3 — Cauda’s anterior pole splits into a dorsoanterior (executive) and a ventroanterior (affective-chemosensory) subdivision. Cauda’s bipartite-plus-transitional split reads as the coarse version of that (their own hierarchical analysis, Chang note, showed “modest support for the tripartite division”); the 10-ROI resolution and a-priori k=2 are what kept the anterior pole undivided here. Read as a refinement, not a disagreement.

The two networks

• Ventral-anterior insula → a limbic / salience pattern. Seeds in the anterior short gyri (ROIs 1, 2, 4, 5, 8) correlate with rostral anterior cingulate (rACC), middle and inferior frontal gyri, temporoparietal junction, cuneus/precuneus and superior temporal gyri. The authors read this as “primarily related to limbic regions which play a role in emotional aspects” — the affective, interoceptive pole.
• Dorsal-posterior insula → a sensorimotor pattern. Seeds in the long insular gyrus (ROIs 3, 7, 10) correlate with premotor, supplementary-motor and sensorimotor cortex, mid-posterior cingulate, superior temporal, lingual and cerebellar cortex — “a role for the insula in sensorimotor integration.”
• A transitional zone. ROIs 6 and 9 (middle short gyrus) show an intermediate profile that the clustering places between the two networks — the functional correlate of the dysgranular mid-insula.

The two networks are not merely distinct but anticorrelated: the pattern negatively correlated with the anterior insula resembles the pattern positively correlated with the posterior insula, and vice versa. In a sample subject the anterior- and posterior-insula time courses correlate r = -0.47, with maximal cross-coherence at ~0.055 Hz. The activation of one network’s structures went with deactivation of the other’s — “reinforcing the idea that the anterior and posterior portions of the insula subserve different functions, and are connected to different networks that operate independently of one another.”

Why the convergence matters for this wiki

The Insular cortex page states the posterior→mid→anterior gradient as established fact and then hangs a dozen studies on it. Cauda et al. supply the resting-state leg of that establishment. Read against the wiki’s other insula evidence:

• It precedes and predicts the task-based dissociations. Haruki & Ogawa (2023) find the right dorsal anterior insula preferring cardiac (arousal) attention and the mid insula coding organ identity; Farb (2013) finds a graded posterior→anterior attentional tuning field. Both are task studies recovering, more finely, the ventral-anterior/dorsal-posterior functional axis Cauda et al. had already drawn at rest.
• It gives the Salience network its intrinsic-connectivity grounding at the insular end. Seeley et al. (2007) defined the salience network from an anterior-insula/ACC co-activation; Cauda et al. show the ventral-anterior insula specifically is the seed that recovers the rACC-anchored limbic pattern, while the dorsal-posterior insula belongs to a separate sensorimotor network — a within-insula dissociation the salience-network literature often glosses.
• It anticipates Chen et al.’s (2021) caution that the salience network is really two networks (a homologous ventral/affective division and a human-elaborated dorsal/attentional one). Cauda et al.’s split is drawn along the insula rather than across the whole network, but it is the same lesson: “the insula” and “the salience network” are each internally divided, and lumping them loses the structure.

The lateralization, and the right-insula arousal reading

The bipartition is partially lateralized, and the pattern lines up with claims elsewhere in the wiki. The salience/anterior network connects more strongly to rACC on the right and to prefrontal cortex on the left; the posterior network connects more strongly to superior temporal and occipital cortex on the right. More pointedly, the right anterior ROIs are preferentially connected to brainstem, pons, and right thalamus, as well as right dlPFC, rACC and right supramarginal gyrus — which the authors read (with Craig 2009; Sridharan et al. 2008; Nelson et al. 2010) as support for the right insula as a pivotal node in arousal and attentional systems.

This is the resting-state counterpart to two things this wiki holds. It fits Craig’s forebrain-emotional-asymmetry hypothesis (right AIC ↔ sympathetic/arousal) — arrived at from connectivity rather than autonomic argument — though, like every correlational result, it cannot test the causal asymmetry claim. And it prefigures the Haruki deflation that the right anterior insula may be an arousal region rather than a general interoceptive one: a structure wired preferentially to brainstem, pons and thalamus (the arousal machinery) is exactly what that later reading predicts.

What it says, and does not say, about prediction

Cauda et al. is a 2010-vintage salience/sensorimotor-network paper; it predates the predictive-coding reframing and takes no position on inferential?. But two of its features bear on that debate as later pages tell it. The anticorrelation of the anterior and posterior insular networks is the resting-state substrate that predictive accounts partition by computational role (anterior = descending prediction, posterior = ascending error, per the EPIC model on Insular cortex) — Cauda et al. show the two ends are functionally opposed at rest without saying why. And the rostroventral→caudodorsal subcortical decorrelation gradient (strong basal-ganglia/thalamus coupling anteriorly, weakening posteriorly) is a connectivity fact any hierarchical model of the insula has to accommodate. The paper offers the anatomy; the computational reading is imported later.

Held at the resting-state discount

Everything here is correlational and cross-sectional — the standing caveat on Resting-state functional connectivity. Three specifics for this study. The 5×5×5 mm seed ROIs exceed cortical thickness, so anterior-insula signal may be partly opercular or subcortical; the authors re-run with resized and repositioned ROIs and report stable maps, but do not eliminate the concern. The sample is age-heterogeneous (mean ~53, SD 19), so this is not a young-adult norm, though age was uncorrelated with motion. And the anticorrelated networks depend on global-signal regression, whose validity for revealing genuine anticorrelation was actively disputed when the paper was written (Fox 2009; Weissenbacher 2009) — the authors flag this and lean on the positive parcellation, which the anticorrelation only confirms. What survives all three caveats is the robust, thrice-replicated positive finding: the resting insula is functionally bipartite along its ventral-anterior/dorsal-posterior axis, matching the anatomy the rest of the wiki assumes.