Node Detection Using High-Dimensional Fuzzy Parcellation Applied to the Insular Cortex

Study at a glance

• Authors: Ugo Vercelli, Matteo Diano, Tommaso Costa, Andrea Nani, Sergio Duca, Giuliano Geminiani, Alessandro Vercelli, Franco Cauda
• Year: 2015
• Journal: Neural Plasticity
• DOI: 10.1155/2015/810183
• N: 20
• Design: Resting-state fMRI (1.5T, 200 volumes) in 20 healthy adults + an 18-subject replication group; connectivity-based parcellation of the insular surface by high-dimensional fuzzy c-mean clustering (time courses concatenated across subjects), with 3 mm spherical ROIs placed at cluster maxima and per-ROI functional connectivity mapped by a dual-regression GLM

Key findings

• A ‘one-step’ node-detection method: fuzzy c-mean clustering both parcellates the insula and detects borders in a single pass, because border voxels have time courses with mixed (nonunivocally attributable) characteristics of adjacent clusters. A fixed 20% fuzziness coefficient classes 20% of voxels as borders; 3 mm spherical ROIs are then dropped at each cluster’s local maximum of homogeneity to serve as network nodes.
• The silhouette validity index selected 12 clusters per insula. Grouping the resulting ROIs into an anterior and a posterior bloc reproduced Cauda et al. (2011): an anterior ventral cingulo-frontoparietal ‘salience’/limbic pattern and a posterior dorsal sensorimotor pattern.
• High-dimensional (12-node) parcellation reveals structure the two-cluster model hides: once the shared anterior/posterior variance is regressed out (dual regression across all 12 ROIs), the insular nodes couple to several distinct networks — default-mode, sensorimotor, and parts of the dorsal attentional network — read as ‘echoes’ (Leech et al. 2012) of whole-brain networks nested within the two main insular patterns, i.e. a hierarchical subparcellation.
• Validates a dorsal vs ventral anterior-insula (dFP/vFP) frontoparietal split (Touroutoglou et al. 2012): a dorsal frontoparietal node centred on dorsolateral/dorsomedial PFC and mid-dorsal cingulate (top-down/bottom-up salience integration) vs a ventral node linked to anterior cingulate, ventral PFC and TPJ (emotional salience + integration of bodily feelings). A cluster just anterior to and between them couples to the default-mode network, overlapping the agranular area of Mesulam & Mufson (1982).
• Node detection was reliable: applied to an independent 18-subject replication group, 17 of 24 ROIs were almost overlapping and the remainder displaced by only a few millimetres. Interhemispheric lateralization recurred, more pronounced in the posterior insula, which showed smaller, less homogeneous clusters than the anterior insula.

Limitations

• Correlational, cross-sectional resting-state fMRI at 1.5T (4.5 mm slices, 200 volumes) — cannot establish direction or say what the network labels compute; the standing resting-state discount.
• Clustering is performed on time courses concatenated across all subjects (a ‘fixed-effect’ step chosen to raise clustering dimensionality); between-subject variance is only reintroduced afterwards at the dual-regression stage, so the parcellation itself is a group-common template rather than a subject-level result.
• Both the number of clusters and the 20% fuzziness coefficient are only semi-principled: the fuzziness fraction is ‘an arbitrary parameter’ set by convention, and although silhouette analysis chose k=12, Kelly et al. (2012) report insular parcellation reliability drops ~50% beyond 12 clusters per side — so 12 sits at the edge of where the method is trustworthy.
• No bodily signal is measured and interoception is not tested — like the sibling parcellation papers, the insula is treated purely as a connectivity object, so its interoceptive relevance is inferred from the recovered anatomy, not shown.

Connections

• Methods: Resting-state functional connectivity
• Concepts: Insular cortex, Salience network, Default mode network, Anterior cingulate cortex (ACC)
• Debates: Is interoception domain-general or organ-specific?, inferential?
• Related studies: Cauda et al. (2011), Deen et al. (2011), Chang et al. (2013), Gong et al. (2015), Craig (2009), Haruki & Ogawa (2023), Brooks et al. (2005)

The fourth member of the wiki’s insula-parcellation family, and the one that pushes the resolution highest. Cauda et al. (2011) split the resting insula into two networks (anterior salience/limbic, posterior sensorimotor); Deen et al. (2011) and Chang et al. (2013) resolved the anterior end into a dorsal-executive and a ventral-affective subdivision (k=3). Vercelli et al. go to k=12 nodes per insula and ask what the two-cluster model hides. This is the same Turin group — Franco Cauda is senior author on both this and the 2011 paper, Alessandro Vercelli on both — so read it as that lab returning to its own bipartition with a higher-dimensional lens rather than as an independent replication. (First author Ugo Vercelli is distinct from senior author Alessandro Vercelli.)

The method: parcellation and border detection in one pass

The paper’s advertised contribution is methodological — a “one-step” node-detection procedure built on the fuzzy c-mean clustering algorithm (see Resting-state functional connectivity). Ordinary (“hard”) clustering assigns each voxel to exactly one parcel; fuzzy clustering assigns a graded membership to every cluster, so a voxel near a border, whose time course is a mixture of two neighbours’ patterns, is left nonunivocally attributed. Setting a fuzziness coefficient of 20% designates 20% of voxels as border voxels, and the borders fall out of the same computation that defines the parcels — no separate edge-detection step. A 3 mm spherical ROI is then placed at each cluster’s local maximum of homogeneity to act as a network node. The stated aim is a simpler way to generate the ROIs that graph-theoretical and seed-based connectivity analyses need.

Two design choices are worth flagging because they bound the result. First, to reach 12 clusters the authors concatenate time courses across all subjects before clustering — a “fixed-effect” step that buys clustering dimensionality (a single 6-minute run has too few time points to support 12 reliable clusters at the subject level) at the price of making the parcellation a group-common template; subject variance is only reintroduced afterward, at the dual-regression connectivity stage. Second, both the fuzziness fraction (20%) and the cluster count are semi-arbitrary: silhouette analysis chose k=12, but the authors cite Kelly et al. (2012) that insular parcellation reliability drops ~50% past 12 clusters, so k=12 sits at the ceiling of where the method is trustworthy.

What the high-dimensional view adds

At the two-cluster level the result is a replication of Cauda 2011: grouping the 12 nodes into an anterior and a posterior bloc recovers the anterior ventral cingulo-frontoparietal “salience” pattern and the posterior dorsal sensorimotor pattern. The new information appears when the shared anterior/posterior variance is regressed out (the dual-regression GLM includes all 12 ROIs simultaneously, so each node’s map is what is left after the others are partialled out). The individual nodes then couple to several distinct whole-brain networks — default-mode, sensorimotor, and parts of the dorsal attentional network — that the two dominant patterns had masked. Following Leech et al. (2012), the authors call these the “echoes” of other networks nested inside the insula, and read them as a hierarchical subparcellation: the two main insular clusters are each built from finer nodes carrying the signatures of the larger networks the insula talks to. This is the same hierarchical-structure claim Chang and Deen make from k=3; Vercelli et al. extend it downward.

The most concrete convergence with the wiki’s other parcellation sources is at the anterior insula. Vercelli et al. validate Touroutoglou et al.’s (2012) split of the anterior insula into two frontoparietal networks — a dorsal (dFP) node on dorsolateral/dorsomedial PFC and mid-dorsal cingulate, doing top-down/bottom-up salience integration, and a ventral (vFP) node on anterior cingulate, ventral PFC and TPJ, doing emotional-salience detection and “the integration of bodily feelings.” That is the same dorsal-executive / ventral-affective seam Chang reached by reverse inference and Deen by k-means-plus-disgust-task — three methods, one anterior subdivision. A further cluster, sitting between and just anterior to the two, couples to the default-mode network and overlaps the agranular area of Mesulam & Mufson (1982), which the authors read as the anterior insula’s pivotal role in reallocating resources between internally- and externally-focused networks (the salience/switching account).

Reliability, and the anterior/posterior asymmetry

Unlike Cauda 2011, this paper carries an explicit reliability check: the whole procedure was re-run on an independent 18-subject replication group, and 17 of 24 ROIs were almost overlapping across the two datasets, the rest displaced by only a few millimetres. The posterior insula was noisier — smaller, less homogeneous clusters and more interhemispheric lateralization than the anterior insula — which the authors read (with Craig) as the posterior insula acting as a heterogeneous “data collector” linked to many networks, versus a more spatially coherent anterior pole. Note this is the opposite framing from the wiki’s usual gradient, where the posterior insula is the orderly primary map (somatotopic, labelled-line) and the anterior the integrative one — but the two are not in conflict: Vercelli et al. are describing connectional heterogeneity (how many networks a region talks to), not the fineness of its body map.

Why it is in this wiki

Like the other three parcellation papers, this is not an interoception study — no bodily signal is recorded, and the insula is treated purely as a connectivity object whose interoceptive significance is imported, not measured. It earns its place for two reasons. It is the highest-resolution confirmation of the hierarchical anterior/posterior organization the Insular cortex page is built on, and independent (if same-lab) support for the anterior insula being itself two subdivisions — the finding Chang and Deen carry. And its “echoes” result — insular nodes wearing the signatures of the default-mode and dorsal-attentional networks — is another data point for the domain-generality reading that this cortex is a general integrator/switch rather than a dedicated body-sense organ, the same deflation reached from task data by Farb and Haruki & Ogawa. Held at the resting-state discount: correlational, 1.5T, a group-template parcellation, and a cluster count at the edge of reliability.