Working with Parcellated Data

In neuromapr: Spatial Null Models and Transforms for Brain Map Comparison

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>",
  fig.width = 6,
  fig.height = 4
)

Most brain imaging analyses eventually need to move between two resolutions: vertex-level (one value per surface point, typically 32k--164k per hemisphere) and parcel-level (one value per brain region, typically 50--400). A cortical thickness map is vertex-level. A table of regional volumes from the Desikan-Killiany atlas is parcel-level. Comparing the two requires converting one to match the other.

neuromapr provides tools for this conversion---aggregating vertices into parcels, mapping parcel values back to vertices, and computing parcel centroids on the cortical surface.

library(neuromapr)

Vertices to parcels

Suppose you have a vertex-level brain map and a parcellation that assigns each vertex to a region. The parcellation is an integer vector where each vertex gets a label, with 0 or NA marking the medial wall (vertices that belong to no region).

set.seed(42)
n_vertices <- 1000
vertex_data <- rnorm(n_vertices)

labels <- sample(
  c(0, 1:10), n_vertices,
  replace = TRUE, prob = c(0.1, rep(0.09, 10))
)

parcel_values <- vertices_to_parcels(vertex_data, labels)
parcel_values

Each parcel gets the mean of its constituent vertices. Medial wall vertices (label 0) are excluded automatically.

The summary function is configurable:

parcel_sd <- vertices_to_parcels(
  vertex_data, labels, summary_func = sd
)
parcel_sd

Parcels back to vertices

The reverse operation paints parcel-level values back onto the vertex mesh. Every vertex gets its parcel's value; medial wall vertices get a fill value.

vertex_filled <- parcels_to_vertices(parcel_values, labels)
head(vertex_filled, 20)

Vertices with label 0 receive NA by default. You can change the fill:

vertex_filled_zero <- parcels_to_vertices(
  parcel_values, labels, fill = 0
)
sum(vertex_filled_zero == 0)

The high-level wrappers

parcellate() and unparcellate() add file-reading convenience on top of the low-level functions. If your data lives on disk as GIFTI files, pass paths directly:

parcel_vals <- parcellate(
  "cortical_thickness.func.gii",
  "aparc.label.gii"
)

The function reads the data file, reads the parcellation labels, validates that they match in length, and calls vertices_to_parcels() underneath.

When working with vectors already in memory, the wrappers still work---they skip the file-reading step:

parcellated <- parcellate(vertex_data, labels)
unparcellated <- unparcellate(parcellated, labels)

all.equal(
  parcels_to_vertices(parcel_values, labels),
  unparcellated
)

The roundtrip is not lossless---parcellation is lossy aggregation. But unparcellate(parcellate(x)) gives you parcel means broadcast back to vertex positions, which is the expected behaviour.

Parcel centroids

Some analyses need to know where each parcel lives in 3D space---for example, to compute a parcel-level distance matrix. get_parcel_centroids() offers three methods for finding the representative point of each parcel.

vertices <- matrix(rnorm(n_vertices * 3), ncol = 3)

Average centroid

Take the mean x, y, z coordinates of all vertices in the parcel. Fast, but the resulting point may not lie on the actual cortical surface.

centroids_avg <- get_parcel_centroids(
  vertices, labels, method = "average"
)
head(centroids_avg)

Surface centroid

Find the vertex closest to the average centroid. The result is always an actual vertex on the surface, which matters when centroids must respect the cortical geometry.

centroids_surf <- get_parcel_centroids(
  vertices, labels, method = "surface"
)
head(centroids_surf)

Geodesic centroid

The most principled approach: find the vertex that minimizes the sum of geodesic (shortest-path-on-surface) distances to all other vertices in the parcel. This requires the face matrix of the surface mesh and the igraph package.

centroids_geo <- get_parcel_centroids(
  vertices, labels,
  method = "geodesic",
  faces = faces
)

Geodesic centroids are computationally expensive---they compute pairwise shortest paths within each parcel. For large parcellations on high-resolution meshes, the surface centroid is a practical approximation.

Building a parcel-level distance matrix

A common use case: you have parcellated data and need a distance matrix for null model generation. Parcel centroids get you there.

parcel_distmat <- as.matrix(dist(centroids_avg))
dim(parcel_distmat)

This distance matrix feeds directly into generate_nulls():

parcel_map <- rnorm(nrow(centroids_avg))
nulls <- generate_nulls(
  parcel_map,
  method = "moran",
  distmat = parcel_distmat,
  n_perm = 100L,
  seed = 1
)
nulls

Parcellation and null models together

The parcel spin methods (baum and cornblath) combine parcellation with spin-based null generation. They need both spherical coordinates and the parcellation vector:

n_lh <- 680
n_rh <- 680
sphere_coords <- list(
  lh = matrix(rnorm(n_lh * 3), ncol = 3),
  rh = matrix(rnorm(n_rh * 3), ncol = 3)
)
parcellation <- rep(1:68, each = 20)
parcel_data <- rnorm(68)

nulls_baum <- null_baum(
  parcel_data, sphere_coords, parcellation,
  n_perm = 50L, seed = 1
)
nulls_baum

The rotation happens at vertex resolution, but the output is parcel-level: one surrogate value per parcel per permutation. This is the right approach when your analysis variable is measured at the parcel level but you want the null model to respect vertex-level spatial structure.

Practical considerations

Parcellation size matters for null models. With 30--50 parcels, distance-based methods like burt2020 and moran work well. With hundreds of parcels (e.g., Schaefer 400), spin-based methods become more attractive because they leverage the full vertex-level spatial structure.

Medial wall handling. All parcellation functions treat label 0 and NA as medial wall. The cornblath null model handles medial wall vertices gracefully during spin-based reassignment. If your parcellation has substantial medial wall coverage, prefer cornblath over baum.

The summary function determines what you are testing. Using mean (the default) tests whether the average value in each region relates to another variable. Using sd would test whether within-region variability relates to it. Choose deliberately.

References

Markello RD, Hansen JY, Liu Z-Q, et al. (2022). neuromaps: structural and functional interpretation of brain maps. Nature Methods, 19, 1472--1480. doi:10.1038/s41592-022-01625-w

neuromapr documentation built on Feb. 27, 2026, 5:08 p.m.