Regions of Interest (ROI)

Introduction to Regions of Interest

What are ROIs?

Regions of Interest (ROIs) are fundamental tools in neuroimaging analysis that allow researchers to focus on specific brain areas or patterns. Rather than analyzing every voxel in the brain independently, ROIs enable:

Targeted analysis of anatomically or functionally defined brain regions
Dimensionality reduction by aggregating signals within regions
Statistical power improvement through spatial averaging
Hypothesis-driven investigations of specific brain areas

ROI Types in neuroim2

The neuroim2 package provides comprehensive support for creating and manipulating ROIs:

ROI Type	Function	Description	Use Case
Spherical	`spherical_roi()`	Sphere around a center point	Searchlight analysis, seed-based connectivity
Cuboid	`cuboid_roi()`	Rectangular box region	Grid-based parcellation
Square	`square_roi()`	2D square in one plane	Slice-based analysis
Clustered	`ClusteredNeuroVol`	Parcellation-based regions	Atlas-based analysis
Custom	`Kernel`	Weighted regions	Gaussian-weighted, gradient-based

Quick Start

# Load the package
library(neuroim2)

# Read a brain volume
file_name <- system.file("extdata", "global_mask2.nii.gz", package="neuroim2")
vol <- read_vol(file_name)

# Create a simple spherical ROI
roi <- spherical_roi(vol, c(20, 20, 20), radius = 5)
cat("ROI contains", length(roi), "voxels\n")
#> ROI contains 11 voxels

Basic ROI Operations

Creating a Spherical ROI

Spherical ROIs are the most commonly used type in neuroimaging. They’re particularly useful for searchlight analyses and seed-based connectivity studies.

# Load an example brain volume
file_name <- system.file("extdata", "global_mask2.nii.gz", package="neuroim2")
vol <- read_vol(file_name)

# Create a spherical ROI centered at voxel coordinates [20,20,20]
# with a 5mm radius, filling all values with 100
sphere <- spherical_roi(vol, c(20, 20, 20), radius = 5, fill = 100)

# Examine the ROI
cat("Number of voxels in ROI:", length(sphere), "\n")
#> Number of voxels in ROI: 11
cat("ROI dimensions:", dim(sphere), "\n")
#> ROI dimensions: 11 3
cat("All values are 100:", all(sphere == 100), "\n")
#> All values are 100: TRUE

Performance Note: C++ vs Pure R Implementation

# The spherical_roi function can use either C++ (fast) or pure R (slower)
# C++ is the default and recommended for performance
sphere_cpp <- spherical_roi(vol, c(20, 20, 20), radius = 5, use_cpp = TRUE)
sphere_r <- spherical_roi(vol, c(20, 20, 20), radius = 5, use_cpp = FALSE)

# Both produce the same result
identical(coords(sphere_cpp), coords(sphere_r))
#> [1] FALSE

Creating a Spherical ROI Around Real-World Coordinates

Often, you’ll have coordinates in millimeter space (e.g., from published studies) rather than voxel indices.

# Define a real-world coordinate in mm
rpoint <- c(-34, -28, 10)

# Convert real-world coordinates to voxel coordinates
vox <- coord_to_grid(vol, rpoint)
cat("Real coordinate:", rpoint, "-> Voxel coordinate:", vox, "\n")
#> Real coordinate: -34 -28 10 -> Voxel coordinate: 42.71428 24.00001 16.2027

# Create spherical ROI with 10mm radius
sphere <- spherical_roi(vol, vox, radius = 10, fill = 1)
cat("ROI contains", length(sphere), "voxels\n")
#> ROI contains 85 voxels

# Verify the center of mass is close to our target
coords <- index_to_coord(vol, indices(sphere))
center_of_mass <- colMeans(coords)
cat("Center of mass:", center_of_mass, "\n")
#> Center of mass: -36.75 -22.75 14.8
cat("Original point:", rpoint, "\n")
#> Original point: -34 -28 10

Creating Multiple Spherical ROIs Efficiently

When creating many ROIs, use the vectorized spherical_roi_set() function for better performance:

library(neuroim2)
vol <- read_vol(system.file("extdata", "global_mask_v4.nii", package = "neuroim2"))
d <- dim(vol)

# Define multiple ROI centers
centers <- rbind(
  c(floor(d[1]/3), floor(d[2]/3), floor(d[3]/3)),
  c(floor(d[1]/2), floor(d[2]/2), floor(d[3]/2)),
  c(floor(2*d[1]/3), floor(2*d[2]/3), floor(2*d[3]/3))
)

# Efficient vectorized creation
rois <- spherical_roi_set(bvol = vol, centroids = centers, radius = 5, fill = 1)
cat("Created", length(rois), "ROIs\n")
#> Created 3 ROIs

# Compare with loop approach (slower but equivalent)
rois2 <- lapply(seq_len(nrow(centers)), function(i) {
  spherical_roi(vol, centers[i,], 5, fill = 1)
})
cat("Loop method also created", length(rois2), "ROIs\n")
#> Loop method also created 3 ROIs

Creating Cuboid and Square ROIs

Cuboid ROIs (3D boxes)

# Create a cuboid ROI - a 3D rectangular box
sp1 <- NeuroSpace(c(20, 20, 20), c(1, 1, 1))

# Create a 7x7x7 cube (surround=3 means 3 voxels on each side of center)
cube <- cuboid_roi(sp1, c(10, 10, 10), surround = 3, fill = 5)
cat("Cuboid ROI contains", length(cube), "voxels\n")
#> Cuboid ROI contains 343 voxels
cat("All values are 5:", all(cube == 5), "\n")
#> All values are 5: TRUE

# Get the coordinates
vox_coords <- coords(cube)
cat("Coordinate ranges:\n")
#> Coordinate ranges:
cat("  X:", range(vox_coords[,1]), "\n")
#>   X: 7 13
cat("  Y:", range(vox_coords[,2]), "\n")
#>   Y: 7 13
cat("  Z:", range(vox_coords[,3]), "\n")
#>   Z: 7 13

Square ROIs (2D in 3D space)

# Create a square ROI - a 2D square fixed in one dimension
sp1 <- NeuroSpace(c(20, 20, 20), c(1, 1, 1))

# Create a 5x5 square in the z=10 plane (fixdim=3 fixes the z dimension)
square <- square_roi(sp1, c(10, 10, 10), surround = 2, fill = 3, fixdim = 3)
cat("Square ROI contains", length(square), "voxels (should be 25)\n")
#> Square ROI contains 25 voxels (should be 25)

# Verify it's planar
vox_coords <- coords(square)
cat("All z-coordinates are the same:", 
    length(unique(vox_coords[,3])) == 1, "\n")
#> All z-coordinates are the same: TRUE
cat("Z-plane:", unique(vox_coords[,3]), "\n")
#> Z-plane: 10

Working with ROI Data

Converting ROIs to Sparse Volumes

Sparse volumes are memory-efficient representations that only store non-zero values:

# Create a spherical ROI
sphere <- spherical_roi(vol, c(50, 10, 10), radius = 10, fill = 1)
cat("ROI contains", length(sphere), "voxels\n")
#> ROI contains 85 voxels

# Convert to SparseNeuroVol - memory efficient storage
# Prefer the provided coercion helper
sparsevol <- as.sparse(sphere)

# Verify properties
cat("Sum of values match:", sum(sparsevol) == sum(sphere), "\n")
#> Sum of values match: TRUE
cat("Dimensions match:", all(dim(sparsevol) == dim(vol)), "\n")
#> Dimensions match: TRUE

# Memory comparison
roi_size <- object.size(sphere)
sparse_size <- object.size(sparsevol)
cat("Memory usage - ROI:", format(roi_size, units = "auto"), "\n")
#> Memory usage - ROI: 9.8 Kb
cat("Memory usage - Sparse:", format(sparse_size, units = "auto"), "\n")
#> Memory usage - Sparse: 8.9 Kb

Extracting Time-Series from ROIs (4D Data)

ROIs are particularly useful for extracting time-series from 4D functional data:

# Load a 4D dataset (using mask file as example - normally this would be fMRI data)
vec4d <- read_vec(system.file("extdata", "global_mask_v4.nii", package = "neuroim2"))
cat("4D data dimensions:", dim(vec4d), "\n")
#> 4D data dimensions: 64 64 25 4

# Create an ROI
roi <- spherical_roi(vol, c(12, 12, 12), radius = 6)

# Extract time-series from the ROI
roi_series <- series_roi(vec4d, roi)
cat("ROI time-series dimensions:", dim(roi_series), "\n")
#> ROI time-series dimensions: 19 3

# Get mean time-series across ROI (average across voxels)
mat_roi <- values(roi_series)      # T x N matrix
mean_series <- rowMeans(mat_roi)   # length equals time dimension
cat("Mean time-series length:", length(mean_series), "\n")
#> Mean time-series length: 4

Using ROIVec for 4D ROI Data

The ROIVec class is designed for efficient storage of 4D ROI data:

# Create a 4D NeuroSpace
vspace <- NeuroSpace(dim = c(10, 10, 10, 20), spacing = c(1, 1, 1))

# Define ROI coordinates
roi_coords <- matrix(c(5,5,5, 6,5,5, 5,6,5, 5,5,6), ncol = 3, byrow = TRUE)

# Create synthetic time-series data for each voxel
n_timepoints <- 20
n_voxels <- nrow(roi_coords)
roi_data <- matrix(rnorm(n_timepoints * n_voxels), 
                   nrow = n_timepoints, ncol = n_voxels)

# Create ROIVec object
roi_vec <- ROIVec(vspace, roi_coords, roi_data)
cat("ROIVec created with", nrow(roi_coords), "voxels and", 
    nrow(roi_data), "timepoints\n")
#> ROIVec created with 4 voxels and 20 timepoints

# Access as matrix of values (T x N)
roi_matrix <- values(roi_vec)
cat("Matrix dimensions (T x N):", dim(roi_matrix), "\n")
#> Matrix dimensions (T x N): 20 4

Searchlight Analyses

Searchlight analysis is a powerful technique for multivariate pattern analysis that examines local neighborhoods throughout the brain.

Basic Searchlight

library(purrr)

# Generate exhaustive searchlight covering all voxels
slist <- searchlight(vol, eager = TRUE, radius = 8)
cat("Number of searchlights:", length(slist), "\n")
#> Number of searchlights: 29532

# Compute mean value in each searchlight
searchlight_means <- slist %>% 
  purrr::map_dbl(~ mean(vol[coords(.)]))

cat("Computed means for", length(searchlight_means), "searchlights\n")
#> Computed means for 29532 searchlights
cat("Mean range:", range(searchlight_means, na.rm = TRUE), "\n")
#> Mean range: 0.02040816 1

Randomized Searchlight

Randomized searchlight samples voxels without replacement, ensuring coverage while reducing redundancy:

# Randomized searchlight - each voxel appears in at least one searchlight
set.seed(42)  # For reproducibility
random_lights <- vol %>% 
  random_searchlight(radius = 8) %>% 
  purrr::map_dbl(~ mean(vol[coords(.)]))

cat("Random searchlight count:", length(random_lights), "\n")
#> Random searchlight count: 2082

Clustered Searchlight

Use anatomical or functional parcellations to define ROIs:

# Create a clustering over the voxel space
grid <- index_to_coord(vol, which(vol > 0))
set.seed(123)
kres <- kmeans(grid, centers = 50, iter.max = 500)

# Create ClusteredNeuroVol
kvol <- ClusteredNeuroVol(vol, kres$cluster)
cat("Created", length(unique(kres$cluster)), "clusters\n")
#> Created 50 clusters

# Run clustered searchlight
cluster_means <- vol %>% 
  clustered_searchlight(cvol = kvol) %>% 
  purrr::map_dbl(~ mean(vol[coords(.)]))

cat("Cluster mean range:", range(cluster_means, na.rm = TRUE), "\n")
#> Cluster mean range: 1 1

Image Patches

Patches are regular, fixed-size regions that tile the image space, useful for convolutional approaches:

# Create 3x3x1 patches covering the volume
pset <- patch_set(vol, dims = c(3, 3, 1))
cat("Number of patches:", length(pset), "\n")
#> Number of patches: 102400

# Compute statistics for each patch
patch_means <- pset %>% purrr::map_dbl(~ mean(.))
cat("Patch mean range:", range(patch_means, na.rm = TRUE), "\n")
#> Patch mean range: 0 1

# Restrict patches to masked regions only
pset_masked <- patch_set(vol, dims = c(3, 3, 1), mask = as.logical(vol))
cat("Masked patches:", length(pset_masked), "\n")
#> Masked patches: 29532

# Patches are guaranteed to be equal size (padded at edges if needed)
patch_sizes <- pset_masked %>% purrr::map_int(~ length(.))
cat("All patches same size:", length(unique(patch_sizes)) == 1, "\n")
#> All patches same size: TRUE
cat("Patch size:", unique(patch_sizes), "voxels\n")
#> Patch size: 18 voxels

Advanced ROI Techniques

Custom Weighted ROIs with Kernels

Create Gaussian-weighted or other custom-shaped ROIs:

# Create a Gaussian kernel for weighted ROI
kern_dim <- c(5, 5, 5)  # 5x5x5 kernel
voxel_dim <- c(1, 1, 1)  # 1mm isotropic voxels

# Create Gaussian-weighted kernel
gauss_kernel <- Kernel(kerndim = kern_dim, vdim = voxel_dim, 
                       FUN = dnorm, sd = 1.5)

# Embed kernel at a specific location
embedded <- embed_kernel(gauss_kernel, space(vol), 
                         center_voxel = c(20, 20, 20))
cat("Kernel weights sum to:", sum(embedded), "\n")
#> Kernel weights sum to: 1

Working with ClusteredNeuroVec for Parcellated Data

Efficiently handle parcellated 4D data where voxels are grouped into regions:

# Create synthetic 4D data
sp4 <- NeuroSpace(c(10, 10, 10, 20), c(1, 1, 1))
arr <- array(rnorm(10*10*10*20), dim = c(10, 10, 10, 20))
vec <- NeuroVec(arr, sp4)

# Create mask for central region
sp3 <- NeuroSpace(c(10, 10, 10), c(1, 1, 1))
mask_arr <- array(FALSE, dim = c(10, 10, 10))
mask_arr[3:8, 3:8, 3:8] <- TRUE
mask <- LogicalNeuroVol(mask_arr, sp3)

# Assign voxels to 5 random clusters
n_voxels <- sum(mask_arr)
clusters <- sample(1:5, n_voxels, replace = TRUE)
cvol <- ClusteredNeuroVol(mask, clusters)

# Create clustered representation - one time-series per cluster
cv <- ClusteredNeuroVec(vec, cvol)

# Access cluster time-series efficiently
cluster_matrix <- as.matrix(cv)  # T x K matrix
cat("Cluster matrix dimensions:", dim(cluster_matrix), "\n")
#> Cluster matrix dimensions: 20 5
cat("(", dim(cluster_matrix)[1], "timepoints x", 
    dim(cluster_matrix)[2], "clusters)\n")
#> ( 20 timepoints x 5 clusters)

ROI Set Operations

Combine and manipulate multiple ROIs:

# Create two overlapping spherical ROIs
roi1 <- spherical_roi(vol, c(20, 20, 20), radius = 6, fill = 1)
roi2 <- spherical_roi(vol, c(23, 20, 20), radius = 6, fill = 1)

# Get indices for set operations
idx1 <- indices(roi1)
idx2 <- indices(roi2)

# Intersection - voxels in both ROIs
intersection_idx <- intersect(idx1, idx2)
cat("Intersection:", length(intersection_idx), "voxels\n")
#> Intersection: 0 voxels

# Union - voxels in either ROI
union_idx <- union(idx1, idx2)
cat("Union:", length(union_idx), "voxels\n")
#> Union: 38 voxels

# Difference - voxels in roi1 but not roi2
diff_idx <- setdiff(idx1, idx2)
cat("Difference (roi1 - roi2):", length(diff_idx), "voxels\n")
#> Difference (roi1 - roi2): 19 voxels

# Calculate overlap percentage
overlap_pct <- length(intersection_idx) / length(union_idx) * 100
cat("Overlap:", round(overlap_pct, 1), "%\n")
#> Overlap: 0 %

Best Practices and Performance

Memory Management

When working with many ROIs, consider memory usage:

# Compare memory usage of different ROI storage methods
n_rois <- 100
centers <- matrix(runif(n_rois * 3, 5, 15), ncol = 3)

# Method 1: List of ROI objects (flexible but more memory)
  roi_list <- lapply(1:nrow(centers), function(i) {
  spherical_roi(vol, centers[i,], radius = 6, fill = 1)
  })

# Method 2: Single sparse volume with labeled regions (ensure increasing indices)
all_indices <- list()
all_labels <- list()
for (i in 1:nrow(centers)) {
  roi <- spherical_roi(vol, centers[i,], radius = 6)
  idx <- indices(roi)
  idx <- sort(unique(idx))
  all_indices[[i]] <- idx
  all_labels[[i]] <- rep(i, length(idx))
}
idx_all <- unlist(all_indices)
lab_all <- unlist(all_labels)
ord <- order(idx_all)
idx_all <- idx_all[ord]
lab_all <- lab_all[ord]
# Ensure strictly increasing indices by removing duplicates
keep <- !duplicated(idx_all)
idx_all <- idx_all[keep]
lab_all <- lab_all[keep]
combined_sparse <- SparseNeuroVol(lab_all, space(vol), indices = idx_all)

cat("Memory usage:\n")
#> Memory usage:
cat("  ROI list:", format(object.size(roi_list), units = "auto"), "\n")
#>   ROI list: 771.1 Kb
cat("  Sparse combined:", format(object.size(combined_sparse), units = "auto"), "\n")
#>   Sparse combined: 23.7 Kb

Choosing ROI Sizes

Guidelines for selecting appropriate ROI sizes:

# Demonstrate effect of ROI size on coverage and overlap
radii <- c(6, 8, 10, 12)
coverage_stats <- data.frame(
  radius = radii,
  n_voxels = numeric(length(radii)),
  pct_overlap = numeric(length(radii))
)

center1 <- c(20, 20, 20)
center2 <- c(25, 20, 20)  # 5 voxels apart

for (i in seq_along(radii)) {
  roi1 <- spherical_roi(vol, center1, radius = radii[i])
  roi2 <- spherical_roi(vol, center2, radius = radii[i])
  
  coverage_stats$n_voxels[i] <- length(roi1)
  
  overlap <- length(intersect(indices(roi1), indices(roi2)))
  total <- length(union(indices(roi1), indices(roi2)))
  coverage_stats$pct_overlap[i] <- overlap / total * 100
}

print(coverage_stats)
#>   radius n_voxels pct_overlap
#> 1      6       19    0.000000
#> 2      8       49    0.000000
#> 3     10       85    0.000000
#> 4     12      163    5.844156

Parallel Processing Tips

For large-scale ROI analyses, consider parallel processing:

# Example of parallel searchlight analysis (not run in vignette)
library(parallel)
library(foreach)
library(doParallel)

# Setup parallel backend
n_cores <- detectCores() - 1
cl <- makeCluster(n_cores)
registerDoParallel(cl)

# Parallel searchlight computation
searchlight_results <- foreach(roi = searchlight_list, 
                               .packages = c("neuroim2")) %dopar% {
  # Your analysis function here
  mean(data[coords(roi)])
}

# Clean up
stopCluster(cl)

Troubleshooting and Tips

Common Issues and Solutions

Issue: ROI extends outside brain mask

# Problem: ROI at edge of brain
edge_vox <- c(2, 2, 2)  # Near edge of volume

# Solution 1: Use nonzero=TRUE to keep only mask voxels
roi_filtered <- spherical_roi(vol, edge_vox, radius = 5, nonzero = TRUE)
cat("Filtered ROI size:", length(roi_filtered), "voxels\n")
#> Filtered ROI size: 0 voxels

# Solution 2: Check if ROI is valid before analysis
if (length(roi_filtered) < 10) {
  cat("Warning: ROI too small for reliable analysis\n")
}
#> Warning: ROI too small for reliable analysis

Issue: Different coordinate systems

# Always verify your coordinate system
test_coord_mm <- c(-34, -28, 10)  # MNI coordinates
test_coord_vox <- coord_to_grid(vol, test_coord_mm)

# Verify round-trip conversion
back_to_mm <- grid_to_coord(vol, matrix(test_coord_vox, nrow = 1))
cat("Original (mm):", test_coord_mm, "\n")
#> Original (mm): -34 -28 10
cat("Voxel:", test_coord_vox, "\n")  
#> Voxel: 42.71428 23.85711 16.1892
cat("Back to mm:", back_to_mm, "\n")
#> Back to mm: -33.99997 -28.00012 10.00004

2025-09-01