Determining the atlas correspondence of Desikan-Killiany-Tourville and Glasser MMP1 atlases across magnetic field strengths

Abstract

Background:

Over sixty-six brain atlases exist to parcellate the brain based on cytoarchitecture, function, and connectivity. Because atlas choice depends on individual study goals and hypotheses, variability in findings contributes to challenges in replication, validation, and reconciling the results across studies. Our goal was to measure the intersection of three commonly used atlases and create a tool to find regional correspondence between the atlases.

New method:

This study used three independent samples of anatomical MRI data acquired with different B0 magnetic field strengths: 1.5 Tesla (T), 3 T, and 7 T. The Desikan-Killiany-Tourville (DKT) and Glasser atlases were used to parcellate the brain. Coefficient-of- variation of regional volumes was measured to evaluate regional variability across subjects in each atlas. DKT and Glasser parcellation correspondence was calculated to answer the shared question of what Glasser regions intersect with a DKT region and vice versa and to investigate consistency of the parcellations in relation to each other across a variety of individuals and image resolutions.

Results:

We found that regional correspondence was consistent across field strengths for the DKT and Glasser parcellations despite showing population variability in volume, age, and sex, and was validated in the Schaefer400 atlas. Parcellation intersection data along with sample code to calculate specific regional correspondence is available.

Comparison with existing methods:

Prior studies have attempted to reconcile multiple atlases, but did not compare voxel- by-voxel on real data.

Conclusion:

This analysis created a tool for researchers and can aid in comparisons with differing atlas choice and variable field strengths.

1. Introduction

The human brain is organized at multiple scales, including topographic maps, lobular (a)symmetry, and conserved sulcal folding patterns. Anatomical organization is central to biological function which is being actively investigated by neuroscientists for potential clinical applications. Parcellation of regions using atlases should reflect the tissue properties to be biologically interpretable (Petersen, Seitzman et al., 2024). Standard anatomically oriented nomenclature is crucial to provide meaning to findings (Gocht and Schumacher, 2023). For instance, the precentral gyrus contains a topographic map of the motor units and can be meaningfully subdivided into distinct sub-regions corresponding to the body, called the homunculus. Finally, while often created at a population level, atlases must conform to match individual anatomy, especially to be useful in surgery and disease research (Toga, Thompson et al., 2001, Evans, Janke et al., 2012).

Many cortical parcellations have been created that satisfy these requirements. As of 2017, at least 66 brain atlases exist (Dickie, Shenkin et al., 2017). Beginning in 1909, Brodmann, considered to be the founder of human brain mapping, published the first parcellation of the entire human cerebral cortex from 2D slices, based on cytoarchitecture (Zilles, 2018). The first attempt to create 3D human brain atlases emerged in the 1950s using the stereotactic system (Nowinski, 2021), and the first digital atlases were produced in 1974, emerging from the need for individualized stereotactic maps (Bertrand, Olivier et al., 1974). In the decades since, researchers have been refining and altering the regional boundaries in the brain based on new evidence from structure and function of specialized areas, increasingly with MRI (Cabezas, Oliver et al., 2011, Evans, Janke et al., 2012, Zilles, 2018). In addition to the type of information used to define regions, the method of atlas generation can also differ. Commonly used methods include group averaging (Glasser, Coalson et al., 2016), voxel-wise averaging (Grabner, Janke et al., 2006), voxel-wise probability (Shattuck, Mirza et al., 2008), Bayesian inference averaging (Van Leemput, Bakkour et al., 2009), and correspondence of MRI regions with postmortem slices (Talairach and Tournoux, 1988).

In research studies, atlas choice is a key step in the analysis and interpretations of the findings (Zalesky, Fornito et al., 2010). Atlas variability may introduce confounds while comparing similar studies with different atlas parcellations. For example, the BOLD signals extracted from the regions and features of major resting state networks differ when regional boundaries differ (Doucet, Lee et al., 2019). Some studies have attempted to standardize the atlases, such as the Neuroparc study (Lawrence, Bridgeford et al., 2021) or the Python package Nilearn (Abraham, Pedregosa et al., 2014). These studies used the Montreal Neurological Institute (MNI) reference brain only, but did not investigate individual variability leaving the question of population heterogeneity unaddressed. Similarly, a 2009 study defines “the brain atlas concordance problem” and introduced quantitative atlas comparison techniques, including a probabilistic interpretation of regional concordance (Bohland, Bokil et al., 2009), but examined atlas correspondence in the single-subject ICBM template brain only. Indeed, individual variability can be pathological, and quantitative atlas overlap techniques may also be valuable in the comparison between more normative atlases to disease-specific atlases (Toga, Thompson et al., 2001). Additionally, it is important for researchers to be able to compare their results with other studies and doing so involves many variables that could potentially affect the regional overlap such as scanner B0 strength. Recent studies have shifted to 7 T data due to higher image resolution and signal-to-noise ratio (SNR) (Calabro, Parr et al., 2024), but 3 T studies are still ongoing and many have been completed using lower field strengths such as 1.5 T. Although acquired with lower resolution, the results from these studies can still be used if we can understand how the scanner strength (and consequently, image resolution) influences these results.

This study takes the atlas correspondence problem to the individual level, and across data acquisition parameters. We studied anatomical MRI data from three field strengths, 1.5 T, 3 T, and 7 T, with voxel resolutions of 1.5 mm, 1 mm, and 0.55 mm, respectively. Our goal is to quantify the regional overlap and variation of parcellations of different atlases across subjects using imaging data acquired for three independent research studies with different spatial resolutions. Based on our analysis, we can further quantify the differences that are common in real data due to anatomical variation from person to person. We compared two atlases with different derivation techniques, number of regions, and motivation behind the parcellation schemes.

We chose the commonly used Desikan-Killiany-Tourville (DKT) atlas because the atlas was derived based on standard neuroanatomical conventions and information from clinicians (Desikan, Segonne et al., 2006). This atlas was developed in 2006 with the goal of producing a reliable automated labeling system for MRI images. The DKT atlas subdivides the human cerebral cortex into 66 gyrally derived regions-of-interest (ROI) based on manual segmentation from 40 subjects. To contrast the DKT, we used the Glasser atlas, derived using multimodal neuroimaging data from the Human Connectome Project (HCP), and was developed 10 years after the DKT. The Glasser atlas was created in a semi-automated pipeline, and parcellated the brain into 180 cortical regions per hemisphere (including the hippocampus), based on a group average of 210 healthy adults. Not only is the Glasser atlas much higher in resolution (number of parcels) compared to the DKT, but it also has regional boundaries based on cortical architecture along with function, connectivity, and topography data (Glasser, Coalson et al., 2016). Many other atlases are available, including classes of atlases based on functional connectivity such as the Schaefer atlas (Schaefer, Kong et al., 2018), however we chose the DKT and the Glasser atlases due to their distinct differences in parcellation resolution and boundary identification methods and validate against the Schaefer atlas.

We hypothesized that the field strength would not affect volumes for DKT atlas, but the field strength would affect the parcels in the Glasser atlas because the latter primarily contains smaller regions defined by connectivity. These volumetric differences will result in atlas correspondence differences between field strength samples. Our goals were to examine the volumetric variation among the groups and test the percent correspondence of atlas parcellations within and between the groups.

2. Methods

This study used three separate anatomical MRI data samples of 30 subjects, and each sample was acquired at different field strengths. Subjects were required to have no personal history of psychiatric disorders, not to have neurological conditions including significant head injury, and not meet criteria for intellectual disability per DSM-IV. Using the anatomical images, the data was quality controlled and parcellated using DKT and Glasser atlases. The variation of regional volumes was measured for each group and this variation was compared between samples. Next the percent correspondence, or the overlap between the two atlases for a given subject, was calculated. Similar to the volumetric comparisons, the variation of the correspondence across the sample was measured and the group variations were compared between samples. A code was designed for the correspondence calculation that was tested on these samples and validated using the Schaefer atlas.

2.1. Imaging data

This analysis was conducted on previously acquired MRI data from three different case-control research studies spanning the last twenty years. These datasets originally consisted of subjects with psychotic disorders and their matched controls. In this study, we chose only the healthy subjects who were enrolled as healthy controls in previous studies. Two sources of data were collected at the Western Psychiatric Institute and Clinic in Pittsburgh, PA, USA. The third data source was from the UK Biobank. We selected 30 subjects from each of these studies/repositories. Our sample size was chosen to be larger than but still representative of sample sizes in highly-cited imaging studies, which on average is about 23–24 subjects in each study (Szucs and Ioannidis, 2020). A sample of 30 is larger than the average size of neuroimaging research studies therefore was sufficient for this comparison of atlas variability.

2.1.1. 1.5 T study

This project was a single site study where the MRI data was acquired on a GE Signa Whole Body 1.5 T scanner. T₁-weighted (T₁w) image voxels were 1.5 mm in thickness, with a 256 × 192 (in-plane: coronal) x 124 (slices) field of view. Images were collected using a 3D spoiled gradient recall (3D-SPGR) sequence (Gilbert, Rosenberg et al., 2001, Prasad, Patel et al., 2004).

2.1.2. 3 T study

These images were sourced from the UK Biobank. The UK Biobank (www.ukbiobank.ac.uk) is a biomedical database containing lifestyle and health data on more than 500,000 participants, including MRI imaging data (Miller, Alfaro-Almagro et al., 2016). Image acquisition parameters in the UK Biobank are published (Littlejohns, Holliday et al., 2020, Smith, Alfaro-Almagro et al., 2022). The UK Biobank is a multi-site study, but all sites used identical Siemens Skyra 3 T scanners with standard Siemens 32-channel receiving coils. T₁w image voxels were 1 mm isotropic, with a 256 × 256 (in-plane: sagittal) x 208 (slices) field of view. Images were collected using a 3D MPRAGE sequence.

2.1.3. 7 T study

The 7 T data came from an early-onset schizophrenia project (R01MH115026–05, PI Prasad) conducted at the University of Pittsburgh that includes imaging data on adolescent onset schizophrenia and demographically matched healthy controls, although we only included controls for this analysis. This is a single-site study using a modified Siemens Magnetom Whole Body 7 T scanner with a custom 64-channel “tic-tac-toe” head coil. T₁w image voxels were 0.55 mm isotropic, with a 390 × 390 (in-pane: axial) x 348 (slices) field of view. Images were collected using a 3D MP2RAGE sequence (Prasad, Muldoon et al., 2023).

2.2. Quality control

Data from all three sources was quality controlled after collection. To ensure high-quality images were used for analysis, we removed scans with ringing or ghosting artifacts. In the case of the 3 T study, we downloaded the data that was quality controlled before release by the UK Biobank (Alfaro-Almagro, Jenkinson et al., 2018). We included 30 randomly selected images with minimal noise/ artifacts using visual inspection. For the 7 T study, we chose the 30 healthy control subjects with the highest internal image quality rating by visual inspection. Out of the 90 subjects across 3 studies included in this analysis, only one subject from 1.5 T dataset failed FreeSurfer’s recon-all step (see next section) and was then replaced. Age and sex were expected to be different between samples due to using previously acquired data for this study which had differing age ranges and sex distributions; the maximum age of the 7 T study was 22 years and the minimum age of the 3 T study was about 40 years. Sex was different because we prioritized choosing high-quality images without artifacts over a sex-matched sample because image quality would be more imperative for parcellation consistency.

2.3. Image processing and parcellation

All T₁w image processing and atlas parcellations were performed using Freesurfer 7 (http://surfer.nmr.mgh.harvard.edu/). FreeSurfer (Fischl, 2012) is a robust and widely used software suite. Image processing steps included removal of non-brain tissue using the watershed method (Ségonne, Dale et al., 2004), segmentation of the subcortical white matter and deep gray matter surfaces, intensity normalization, tessellation of the gray-white matter boundary, automated topology correction, and intensity defined tissue classification. After this pre-processing, FreeSurfer reconstructs the cortical surfaces (as vertices) from the native-space T₁w images (Fischl, Sereno et al., 1999, Desikan, Segonne et al., 2006). For any atlas of interest, given a FreeSurfer label file (.annot file) the atlas labels can be mapped to the cortical surface using mri_surf2surf (Fischl, van der Kouwe et al., 2004) restricting labels to grey matter voxels only and deforming the atlas to individual anatomy. These surface files are then converted back to volumes with mri_aparc2aseg, thus providing a subject-space version of the cortical labeling of gray matter in FreeSurfer native space.

The DKT parcellation is automatically created using the standard recon-all Freesurfer command. The Glasser parcellation was acquired by further processing the Freesurfer output directories using an fs_average version (Mills, 2016) of the Human Connectome Project (HCP) multimodal parcellation version 1 (MMP1) in order to map the atlas to the individual’s native space (Neurolab, 2017). Scripts to perform these steps are on the GitHub (https://github.com/kmrprasad/Atlas_correspondence). The 7 T images were processed using the “high-res” Free-surfer option.

2.4. Quantitative atlas-based calculations

2.4.1. Volumetric analysis

Regional volume was calculated for each region in both atlases for each individual using the voxel labeling provided by each parcellation scheme. The total number of voxels in each parcel was converted to volume measured in cubic millimeters (mm³) in order to estimate regional volumes. Total cortical brain volume (TCBV) in cubic millimeters was calculated as the number of voxels included in the brain mask. White matter volume (WMV), gray matter volume (GMV) and cerebrospinal fluid volume (CSFV) were acquired from the Freesurfer aseg_stats.txt files for each individual from the columns “Cerebral-WhiteMatterVol”, “TotalGrayVol”, and “CSF”, respectively, and was measured in cubic millimeters. Regional volumes were compared across samples for each atlas using a multivariate analysis of covariance (MANCOVA) controlling for age, sex, and TCBV. Coefficient of variation (CoV) of regional volume was measured across the subjects in each sample for each region. The CoV is the ratio of the standard deviation over the mean. A high CoV indicates high variability across the datasets and low CoV indicates relatively less variability between two datasets. The CoV of volumes were compared across field strength samples for each atlas using t-tests. Regions from both the atlases with a volume CoV≥ 2 standard deviations higher than the mean were identified as outliers.

2.4.2. Atlas correspondence

In the 1.5 T and 3 T samples, each subject’s MRI was resampled into the Freesurfer native space (256x256x256). For the 7 T MRI data, we used the high-resolution Freesurfer native space (390x390x390). We compared the parcellations from DKT atlas with Glasser atlas of the same subject’s MRI data from the same scanner. To clarify further, MRI data from 1.5 T was parcellated for DKT and Glasser atlas and the Percent Correspondence (PC) of two regions from the same subject was examined. We did not compare atlases from one field strength with atlas parcellations from another field strength due to differing resolutions. To explain further, two regions that do not spatially correspond at all with each other will have PC= 0. In this study, we only compared atlases within individual subjects so that voxels would exactly align with one another which also preserved individual differences in anatomy.

PC was calculated using one atlas as a reference and another as a target. For example, when DKT is the reference atlas then Glasser atlas is the target and vice versa. PC is a measure of the overlap between one region in the reference atlas with another region in target atlas. PC values range from 0 to 1. A PC= 0 means that the regions do not spatially overlap at all and PC= 1 means the entire reference region corresponds to the single target region. A low PC indicates a small amount of correspondence/overlap between the two regions and a high PC, close to 1, indicates high correspondence/overlap of the regions. The PC of each region in the reference atlas with each region in target atlas was calculated using

$$P{C}_{i,j}=\frac{N_j}{M_i}$$

where i is a region in the reference atlas, j is a region in the target atlas, M is the total number of voxels in region i, and N is the total number of j’s voxels within region i (Fig. 1). When two atlases with unequal parcel sizes are compared, it is expected that when the atlas with smaller parcels is the reference atlas, many PC values are likely to be closer to one.

Fig. 1.

Calculation of Percent Correspondence: Squares represent voxels; all voxels are in the same coordinate system. For each region in the reference atlas regions in the target atlas are spatially compared. In this example the reference is the DKT atlas and the Glasser atlas is the target atlas. Percent Correspondence (PC) is calculated as the percentage of the reference atlas region voxels which correspond to a region in the target atlas. As in step 3, 29 voxels correspond between the DKT region and the Glasser region. Since there are 76 voxels total in the reference (DKT) region, the PC is 29/76 or 38 %. This calculation is repeated for all regions in the target and reference atlas. Because it is uncommon for regional boundaries to correspond exactly to one another, there will be Glasser region voxels which do not correspond to the DKT region voxels, shown in step 3 as red voxels and step 4 as yellow voxels. To find which DKT regions correspond with each Glasser region, the PC calculation will need to be repeated with the reference atlas as the Glasser atlas and the DKT as the target.

Calculating PC_i,j in this way for each possible pair of regions between atlases produces the PC matrix for a subject. Regional boundaries may not directly correspond between one atlas and another, and some target atlas voxels will correspond to a different reference atlas region as shown in Fig. 1 part 3 (red voxels) and 4 (yellow voxels). Therefore, the PC matrix is not symmetric, and to find which DKT regions correspond with each Glasser region, the PC calculation was repeated with the reference atlas as the Glasser atlas and the DKT as the target. This produced two PC matrices for each subject, one 66 × 358 matrix (DKT-reference, Glasser-target) and one 358 × 66 matrix (Glasser-reference, DKT-target). Where correspondence was less than 1 %, we treated it as noise by setting the correspondence to 0 %, such as left and right hemisphere parcels slightly overlapping on the midline. The mean and standard deviation across subjects in each sample was calculated. To compare between field strength groups, we had a 1.5 T averaged PC matrix, 1.5 T standard deviation of PC matrix, 3 T averaged PC matrix, 3 T standard deviation of PC matrix, 7 T averaged PC matrix, and 7 T standard deviation of PC matrix.

To test the similarity across field strengths, the averaged PC matrices were vectorized and plotted and the coefficient of determination adjusted the reported R² values. Coefficient of determination measures model fitness in a regression between two variables whereas CoV measures relative variability within one variable. The regions with the highest and lowest correspondence between DKT and Glasser atlases were identified by calculating the maximum and minimum averaged PC, respectively. We will examine these regions’ correspondence and illustrate where applicable. We will also investigate the regions with high CoV of volume in the same way. In addition to these quantitative examinations, we provide an example of how this data can be used in the research field (https://github.com/kmrprasad/Atlas_correspondence).

2.5. Validation of percent correspondence method

To validate our percent correspondence calculation, we chose the Schaefer 400 atlas, to compare with the DKT and Glasser atlases because it represents another class of atlases, namely functional atlases, with a derivation method that is different from the DKT and Glasser atlases. Whereas the DKT and Glasser atlases rely on histology, neuroanatomical information and multimodal data, functional parcellations such as the Schaefer atlas rely on fMRI-derived intrinsic functional connectivity, using task and resting-state fMRI data to generate gradient-weighted Markov Random Field parcellations (Schaefer, Kong et al., 2018). To create subject-space Schaefer atlas parcellations for validation, first the FreeSurfer annotation files for the Schaefer 2018 400 parcel atlas with the Yeo 7-network naming conventions were downloaded from the public GitHub repository (Schaefer, Kong et al., 2018). Then, as described above, the FreeSurfer command mri_surf2surf was used in conjunction with the subject-level FreeSurfer directories to map the Shaefer annotation files to each subject’s cortical surface, and finally these surfaces were converted to volumes using the mri_aparc2aseg command. We calculated the PC of the Schaefer atlas between the DKT and Glasser atlases separately and analyzed in the same way as the DKT and Glasser comparisons to validate atlas correspondence on a third condition.

3. Results

3.1. Demographics and brain volumes

The final sample consisted of 30 subjects in each field strength group: 1.5 T, 3 T, and 7 T. As expected, groups were significantly different by age and sex distribution (Table 1). TCBV, GMV, WMV, and CSFV did not differ between the groups after controlling for age and sex (Table 1).

Table 1.

Comparison of samples on demographic data, total cortical brain volume (TCBV), gray matter volume (GMV), white matter volume (WMV), and cerebrospinal fluid volume (CSFV) for 1.5 T, 3 T, and 7 T samples (A). All volume measurements are in cubic millimeters (mm³). First, overall demographics were calculated to compare between field strength samples (B). A one-way ANOVA was used to compare age between the groups. Pearson chi-squared test of proportions was used to compare sex. TCBV, GMV, WMV, and CSFV were compared using a univariate analysis of variance with age and sex as covariates. Post-hoc tests were completed to compare each group individually. LSD test was used for age, TCBV, GMV, WMV, and CSFV post-hoc comparisons we are reporting uncorrected p-values. Significance is reported after adjustment for multiple comparisons. Discrepancies between TCBV and the addition of GMV, WMV, CSFV is that GMV includes both cortical and subcortical gray matter which was not included in TCBV.

A.	Sample Characteristic
	1.5 T			3 T		7 T
Age (years)	25.13 ± 6.79			56.37 ± 6.97		19.17 ± 1.51
Sex (M/F)	18/12			10/20		9/21
TCBV	1.22 E 06 ± 1.11 E 05			1.15 E 06 ± 1.09 E 05		1.16 E 06 ± 1.05 E 05
GMV	6.68 E 05 ± 6.08 E 04			5.88 E 05 ± 5.08 E 04		6.73 E 05 ± 5.85 E 04
WMV	4.71 E 05 ± 5.79 E 04			4.74 E 05 ± 5.36 E 04		4.28 E 05 ± 5.08 E 04
CSFV	9.12 E 02 ± 1.56 E 02			1.10 E 03 ± 2.20 E 02		1.00 E 03 ± 2.55 E 02
B.	Sample Comparisons
	Overall			1.5 T vs 3 T	1.5 T vs 7T	3 T vs 7 T
	Test Statistic	df	p-val	Significance	Significance	Significance
Age	F= 370.67	2	< 0.001	< 0.001	< 0.001	< 0.001
Sex	χ2= 6.70	2	0.035	0.04	0.02	0.78
TCBV	F= 0.90	2	0.41	0.73	0.19	0.42
GMV	F= 0.42	2	0.66	0.75	0.42	0.98
WMV	F= 2.19	2	0.12	0.54	0.04	0.19
CSFV	F= 2.31	2	0.10	0.50	0.06	0.88

3.2. Regional volumetric analysis

Our main goal was not to examine volumetric differences but to investigate the volumetric inter-individual variability and identify regions with high variability in each of the field strength samples. For all field strength samples, the largest region in the DKT atlas was the superior frontal gyrus and the smallest was the transverse temporal region. In the Glasser atlas, the primary visual cortex was the largest region and area 52 was the smallest region across all field strengths. Thirty-four DKT regions were significantly different across the samples and 169 Glasser regions were different (Supplemental Table 1,2). In the MANCOVA model for regional volumes, age and sex were not significant for the DKT regions (both p > 0.05) but were significant for the Glasser region (Age: p = 0.04, Sex: p = 0.007). In the test for between subjects effects, out of 358 regions 42 and 28 were significantly different for age and sex, respectively. The CoV of regional volumes in the 7 T sample was 20 % higher than 3 T and about 13 % higher than the 1.5 T samples (both p < 0.001) (Table 2). Spatially, DKT and Glasser atlas regional volume CoV were different across the field strength samples but followed similar patterns (Fig. 2). In the DKT atlas, outliers with high CoV were the left entorhinal cortex in all samples, the right caudal anterior cingulate in both 1.5 T and 3 T samples, the right pars triangularis in the 3 T sample, and the right entorhinal cortex in the 7 T sample. In the Glasser atlas, 12, 11, and 13 regions were outliers with high CoV in 1.5 T, 3 T, and 7 T samples, respectively (See Supplemental Table 3). The right Temporo-Parieto-Occipital Junction (TPOJ) 3 was the only region with high CoV that was common to all field strength samples (Supplemental Table 3).

Table 2.

Coefficient of Variation of Regional Volumes and Sample Comparisons. Average coefficient of variation (CoV) across atlas regions is shown for each atlas at each field strength sample (A). Comparisons were made using a one-way ANOVA and LSD for post-hoc comparisons (B).

A. Coefficient of Variation of Regional Volumes				B. Coefficient of Variation Comparisons
B0 Strength	1.5 T	3 T	7 T	Overall			1.5 T vs 3 T	1.5 T vs 7 T	3 T vs 7 T
	Mean ± sth	Mean ± sth	Mean ± sth	F	df	p	Sig.	Sig.	Sig.
DKT CoV	0.13 ± 0.04	0.12 ± 0.04	0.15 ± 0.05	8.38	2	< 0.001	0.30	0.004	< 0.001
Glasser CoV	0.18 ± 0.05	0.17 ± 0.05	0.23 ± 0.07	108.18	2	< 0.001	0.11	< 0.001	< 0.001

Fig. 2.

Volume Coefficient of Variation: Regional volume coefficient of variation (CoV) across subjects shown spatially for 1.5 T (left), 3 T (middle), and 7 T (right) samples. A. the Glasser atlas and B. the DKT atlas. Red indicates high CoV and dark blue indicates low CoV in each region.

3.3. Atlas correspondence

3.3.1. Consistency of percent overlap across field strengths

The correspondence for each of the 358 cortical Glasser regions to each of the 66 DKT regions resulted in 23,628 correspondence measures regardless of which atlas is used as target/reference. Since many regional pairs have no exact correspondence between atlases, such correspondence pairs were removed to ensure results were not skewed towards 0 % correspondence. Out of a total of 23,628 possible regional correspondence pairs (points in the PC scatter plots at 0,0) only 1,214 data points remained after eliminating non-corresponding node pairs. However, all regions in both atlases are represented in our comparison and correspondence analysis. Similarity of percent overlap by field strength comparison is shown in the plots below where axes represent the percent overlap in the given field strengths (Fig. 3). Scatter plot points represent the group level average for each possible pair of parcels comparisons between atlases.

Fig. 3.

Similarity of percent correspondence by field strength comparison: Axes represent the percent overlap in the given field strengths. Scatter plot points represent the group level average for each possible pair of parcels comparisons between atlases. The adjusted R² values are the coefficient of determination. Scatter plot points closer to the x = y line indicates tighter agreement between field strengths for that inter-atlas region-pair and points closer to one axis indicate a greater overlap was measured for the two regions in the axis (study field strength) the point is closer to. There is high agreement of atlas overlap regardless of reference atlas choice and field strength comparison. In the bottom row corresponding Glasser to DKT, 47 % of the 1214 points showed < 5 % correspondence and in the top row 71 % for DKT as Glasser. With the Glasser atlas as the reference and the DKT atlas as the target (D-F), there were 11 % of the points with > 95 % correspondence indicating that Glasser regions have high correspondence to one DKT region. A) Relative to the DKT atlas, where the DKT atlas is the reference and examining the percentage overlap for each Glasser atlas region for the 1.5 T study and the 3 T. B) Comparing 3 T and 7 T relative to the DKT atlas. C) Comparing 1.5 T and 7 T relative to the DKT atlas. D) Comparing 1.5 T and 3 T relative to the Glasser atlas, where the percent correspondence of each DKT parcel was calculated for each Glasser parcel. E) Comparing 3 T and 7 T relative to the Glasser atlas. F) Comparing 1.5 T and 7 T relative to the Glasser atlas.

All vectorized group-level PC matrices were stable across field strengths with each comparison showing coefficient of determination R² > 0.94 (Fig. 3). Regardless of field strength or reference atlas, regional percent correspondence was always consistent. This was supported by the largest average absolute error terms for region pairs being on average 0.031 with standard deviation of 0.052 for the 3 T vs 7 T comparison for DKT to Glasser (Supplemental Table 4) and distributions of absolute difference in percent correspondence between field strength groups being skewed towards zero error (Supplemental Figure 1).

This plot also shows that one DKT region (pericalcarine) in each hemisphere (shown in the upper right corner of plots A-C in Fig. 3 and corresponds to two overlapping points for the left and right hemispheres) has almost 100 % correspondence with one Glasser region (V1-primary visual cortex). Ninety-nine percent (1,210/1,214) of the other PC values are below 50 %, meaning many Glasser regions are needed to fully encompass one DKT region (Fig. 3A–C). Regardless of the nature of each region’s overlap, atlas-to-atlas regional overlap differences were not found between scanner field strengths.

Because none of these correspondences were found to be appreciably different between field strengths, and because the 7 T data offers the highest image resolution, subsequent figures with MRI images use a representative individual MRI scan from the 7 T study. As described in the methods section, we have focused the remainder of this analysis on demonstrating correspondence results for selected and illustrative regions.

3.3.2. High correspondence and directionality

This first comparison was made to illustrate the concept of directionality of the atlas correspondence, using the DKT region with the highest correspondence to a single Glasser region across subjects because it is important to carefully choose the directionality of the question. For example, in the DKT atlas, the pericalcarine cortex had high correspondence (99.7 %) with the Primary Visual Cortex (V1) in the Glasser atlas (Fig. 4A). It may seem from these findings that the DKT region pericalcarine and the Glasser region V1 overlap exactly, but that is not the case. If the directionality of the question is switched, the Glasser region V1 is 34 % pericalcarine, 27 % lingual, 24 % lateral occipital, and 14 % cuneus regions of the DKT atlas (Fig. 4B).

Fig. 4.

High Correspondence and Directionality: The DKT region with the highest correspondence to a single Glasser region is shown for one subject. A: DKT (blue) as the reference and Glasser (red) as the target atlas. DKT region Pericalcarine has 100 % correspondence with Glasser region V1. B: Glasser (red) as the reference and DKT (shades of blue) as the target atlas. Glasser region V1 corresponds to pericalcarine and three other DKT regions. Despite V1 explaining 100 % of the pericalcarine, the pericalcarine only explains only 34 % of V1; this is a visual representation of atlas correspondence directionality. This visual inspection may not reflect percentages due to the 3D nature of parcellations.

3.3.3. Low correspondence

A second comparison consisting of the DKT region with the lowest correspondence was made, considering only the DKT as the reference. The DKT region with the lowest correspondence is the Superior Frontal region. This is the largest region in the DKT atlas and has 21 Glasser regions that correspond with its voxels. The PC was low (≈10 %) with the maximum correspondence from one region, Area 9 medial. The other regions in descending order of PC were Area 6 m anterior, Area 10d, Superior Frontal Language Area, Area 9 anterior, Supplementary and Cingulate Eye Field, Area 8B medial, Area 8B lateral, Area 9 posterior, Area posterior 10p, Area 6 anterior, Area dorsal 32, Area 8Ad, superior 6–8 Transitional Area, Area 6 mp, Area p32 prime, Polar 10p, dorsal Area 24d, Area 9–46d, Area 10 v, and ventral Area 24d. The superior frontal expands across 5 Glasser cortical sections: Anterior Cingulate and Medial Prefrontal Cortex, Dorsolateral Prefrontal Cortex, Orbital and Polar Frontal Cortex, Paracentral Lobular and Mid Cingulate Cortex, and Premotor Cortex Fig. 5.

Fig. 5.

Low Correspondence with DKT as reference: A. A reference region from the DKT atlas is shown, the superior frontal cortex. This region has the lowest correspondence to any one Glasser region, only 10 % correspondence with its most explanatory region. B. Twenty-one Glasser atlas (target) regions correspond to the superior frontal cortex. Due to the 3D nature of cortical areas not all parcels are visible in any given slice and this visual representation only depicts 12 regions while 9 are not in view.

3.3.4. Answering unique research questions

A third exemplary correspondence assessment consisting of a region with high volume CoV was completed to evaluate whether a region with high CoV of volume across subjects would also vary in atlas correspondence across subjects. To address this, we examined the entorhinal cortex in the DKT atlas since it had the highest CoV of volume in all field strengths (Supplemental Table 3). Across all field strengths, it corresponded with about 50 % perirhinal ectorhinal cortex (PeEC) and 50 % entorhinal on the Glasser atlas on average across subjects with a relatively narrow dispersion of PC (4 % for 1.5 T, 7 % for 3 T, and 11 % for 7 T).

Finally, a fourth comparison motivated by a unique research question was investigated. We investigated DKT parahippocampal region to find which Glasser regions correspond with our past results in this DKT region (Prasad, Rohm et al., 2004, Prasad, Muldoon et al., 2023). We found that on average the parahippocampal DKT region corresponds with Parahippocampal Area 1 (40.7 %), Presubiculum (21.6 %), PeEc (12.9 %), Parahippocampal Area 2 (12.2 %), and Entorhinal Cortex (10.4 %).

3.4. Validation of percent correspondence method

3.4.1. DKT and Schaefer correspondence

For the validation of PC in another atlas, namely the Schaefer atlas, all vectorized group-level DKT and Schaefer PC matrices were stable across field strengths with each comparison showing coefficient of determination R² > 0.91 (Fig. 6). Regardless of field strength or reference atlas, regional PC within subjects was always consistent. This was supported by the largest average absolute error term for region pairs being a mean of 0.032 with standard deviation of 0.045 for the 3 T vs 7 T comparison for DKT to Schaefer (Supplemental Table 4), and distributions of absolute difference in percent correspondence between field strength groups skewed towards zero error (Supplemental Figure 2).

Fig. 6.

Similarity of percent correspondence of DKT and Schaefer atlas by field strength comparison: Axes represent the percent overlap in the given field strengths. Scatter plot points represent the group level average for each possible pair of parcels comparisons between atlases. The adjusted R² values are the coefficient of determination. Scatter plot points closer to the x = y line indicates tighter agreement between field strengths for that inter-atlas region-pair and points closer to one axis indicate a greater overlap was measured for the two regions in the axis (study field strength) the point is closer to. There is high agreement of atlas overlap regardless of reference atlas choice and field strength comparison. In A-C, DKT-reference and Schaefer- target. In D-F, Schaefer-reference and DKT- target. A) Relative to the DKT atlas, where the DKT atlas is the reference and examining the percentage overlap for each Schaefer atlas region for the 1.5 T study and the 3 T. B) Comparing 3 T and 7 T relative to the DKT atlas. C) Comparing 1.5 T and 7 T relative to the DKT atlas. D) Comparing 1.5 T and 3 T relative to the Schaefer atlas, where the percent correspondence of each DKT parcel was calculated for each Schaefer parcel. E) Comparing 3 T and 7 T relative to the Schaefer atlas. F) Comparing 1.5 T and 7 T relative to the Schaefer atlas.

3.4.2. Glasser and Schaefer correspondence

For the Glasser and Schaefer atlas comparison, all vectorized group-level PC matrices were stable across field strengths with each comparison showing coefficient of determination R² > 0.93 (Fig. 7). Regardless of field strength or reference atlas, regional PC was always consistent. This was supported by the largest average absolute error term for region pairs of 0.032 with standard deviation of 0.043 for the 1.5 T vs 7 T comparison for Schaefer to Glasser (Supplemental Table 4), and distributions of absolute difference in percent correspondence between field strength groups skewed towards zero error (Supplemental Figure 3).

Fig. 7.

Similarity of percent correspondence of Glasser and Schaefer atlas by field strength comparison: Axes represent the percent overlap in the given field strengths. Scatter plot points represent the group level average for each possible pair of parcels comparisons between atlases. The adjusted R² values are the coefficient of determination. Scatter plot points closer to the x = y line indicates tighter agreement between field strengths for that inter-atlas region-pair and points closer to one axis indicate a greater overlap was measured for the two regions in the axis (study field strength) the point is closer to. There is high agreement of atlas overlap regardless of reference atlas choice and field strength comparison. In A-C, Glasser-reference and Schaefer- target. In D-F, Schaefer-reference and Glasser-target. A) Relative to the Glasser atlas, where the Glasser atlas is the reference and examining the percentage overlap for each Schaefer atlas region for the 1.5 T study and the 3 T. B) Comparing 3 T and 7 T relative to the Glasser atlas. C) Comparing 1.5 T and 7 T relative to the Glasser atlas. D) Comparing 1.5 T and 3 T relative to the Schaefer atlas, where the percent correspondence of each DKT parcel was calculated for each Schaefer parcel. E) Comparing 3 T and 7 T relative to the Schaefer atlas. F) Comparing 1.5 T and 7 T relative to the Schaefer atlas.

4. Discussion

In neuroimaging, atlases are used to partition the brain into meaningful regions. Because there are many atlases to choose from when designing an experiment, replicating and comparing results between studies using different atlases can be challenging. An important question that we aimed to answer pertains to what region/regions from one atlas correspond to regions from a different atlas at the level of the individual subject, and whether these properties are susceptible to variables common to all MRI studies, such as participant age, sex, and scanner field strength. Our study focused on two widely used digital atlases, the DKT and Glasser MMP1, and their correspondence with each other using three samples of MRI data acquired on different field strengths, and consequently different image resolutions. We also included a third atlas, the Schaefer400, as validation of the consistency of the percent correspondence results. The goal of our study was not to determine the exact nature of the physical effect of the field strength on atlas overlap, but to assess if there were any identifiable effects on atlas overlap associated with the field strength while controlling for other known confounds. Based on the findings above, individual atlas mapping is supported for a wide range of atlases sizes, from broad to granular parcellations in both high- and low-resolution 3D images. While high resolution images (0.55 mm at 7 T) yield significantly higher variability measurements of individual morphometry (Table 1) the pattern of volumetric variability is spatially localized in the same manner across conditions (Fig. 2) and atlas correspondence is robust to changes in population demographics and imaging resolution (Figs. 3, 6 and 7).

Many regional volume differences were found within each atlas between groups. However, the group differences that we observed for most regional volumes within a particular atlas and between groups were not directly attributable to field strength due to age, sex, and TCBV. There were significant effects of age and sex between groups, but despite differences in scanner hardware and demographic variability, atlas correspondence was found to be highly robust across DKT and Glasser atlas comparisons. Percent correspondence with the Schaefer400 atlas also proved to be consistent across field strengths in both the DKT and Glasser validation comparisons. Because the CoV of regional volumes across subjects in each group was significantly higher in the 7 T sample, it is possible that the better voxel resolution leads to more accurate volume estimates, like the so-called “coastline paradox," where finer approximation leads to amplified shape complexity and therefore a more accurate estimation of volumes (Gardiner, Behnsen et al., 2018). In the context of imaging samples, it will produce greater amounts of morphometric inter-individual variability. This is reflected in the correspondence investigation of the high volume CoV reference region, the DKT entorhinal. With DKT entorhinal as reference, the Glasser PeEc and Glasser entorhinal are the two corresponding target regions. Here the dispersion of PC increased nominally as field strength increased: 4 % for 1.5 T, 7 % for 3 T, and 11 % for 7 T. These findings suggest that the contours of the regions and the relative sizes of regions may vary marginally according to field strength, but brain atlases are robust to scanner field strength and demographic composition.

To perform these comparisons, we have developed a MATLAB function available to researchers to compare the correspondence between any two atlases. Prior studies have attempted to reconcile multiple atlases and compare them, but did not compare voxel by voxel imaging data or investigate atlas correspondence unlike our study. (Dickie, Shenkin et al., 2017, Lawrence, Bridgeford et al., 2021). One of the earliest studies on the atlas concordance problem concluded that spatial overlap was sufficiently complex that quantitative measures are necessary (Bohland, Bokil et al., 2009). However, this study only relied on a single standard brain image to quantify overlap. Additionally, our findings consider new and detailed atlases that have been developed since this earlier study (Glasser, Coalson et al., 2016, Schaefer, Kong et al., 2018). To our knowledge no studies have examined the relationship between atlas correspondence and image resolution/field strength. Due to this gap in the literature, we used subject data acquired at three field strengths to report atlas correspondence for diverse MRI protocols.

More recently available atlases also have regional boundaries based on novel considerations. For example, the Glasser MMP1 boundaries were based on cytoarchitecture and multimodal MRI data, thus the atlas itself carries some information about histology, regional functional, and structural connectivity (Glasser, Coalson et al., 2016). The Schaefer400 atlas, also included in this study, was published in 2017 and utilizes fMRI-derived intrinsic functional connectivity to generate gradient-weighted Markov Random Field parcellations (Schaefer, Kong et al., 2018). Our tool can be used to correspond results from one atlas to another to reconcile the results from different atlases. A previous analysis found significant differences in the parahippocampal region of the DKT atlas (Prasad, Rohm et al., 2004). The Glasser atlas was not available at the time of that original study, and we wished to extrapolate these earlier findings onto this newer atlas. After investigation, we found that the parahippocampal areas of the DKT atlas corresponds to five Glasser regions: parahippocampal area 1, presubiculum, PeEc, parahippocampal area 2, and entorhinal cortex. Thus, if future analyses using the Glasser atlas report results in these areas, it is possible that the previous findings would be replicated. Similarly, our recent analysis investigating structural covariance networks was completed using 1.5 T data (Lewis, Santini et al., 2023), but we are currently acquiring a similar sample using 7 T MRI. If we would want to replicate this analysis on the 7 T sample, it would be advantageous to know the volumetric and correspondence differences between field strength samples. Additionally, one could also use our atlas overlap analysis tools to find more information about study-specific regions. For instance, in machine learning, it is advantageous to use an atlas with fewer regions (Lei, Qin et al., 2022), but valuable information is lost when using a less detailed atlas (Lewis, Jiang et al., 2025).

To overcome this, one could report findings in the DKT atlas and extrapolate them into the more detailed Glasser atlas. For example, when the DKT-defined entorhinal cortex is a region of interest, extrapolation of entorhinal cortex to the Glasser atlas could provide more information on the pathophysiological significance and connectivity of the entorhinal cortex with other regions (Prasad, Patel et al., 2004, Alexander, Loh et al., 2019). The DKT entorhinal cortex corresponds with Glasser entorhinal and Glasser PeEc, both in the medial temporal cortex Glasser-defined section. The Glasser entorhinal has more myelin and differs in functional connectivity relative to the PeEc. These two regions may constitute meaningful subdivisions of the larger DKT-defined entorhinal cortex. The PeEc is a newly defined region by the Glasser atlas and is comprised of what would typically be considered the peri-entorhinal cortex and the ectorhinal cortex (Glasser, Coalson et al., 2016). Peri-entorhinal cortex has different functional connectivity than its neighbors and, based on the HCP data, may be the site of the anterior temporal face patch (Glasser, Coalson et al., 2016). Although the information from the DKT atlas is useful for certain applications, the Glasser atlas adds more useful information based on MRI connectivity and task fMRI data.

A limitation of this study is the sample size and the differences in age and sex. A larger and age/sex matched sample could give more consistent results, but this is not the nature of scientific studies. The goal of this analysis was not to represent population variance but instead attempted to measure variance in common neuroimaging research studies. A review manuscript investigating sample sizes in highly cited neuroimaging manuscripts reported 23–24 subjects per study on average (Szucs and Ioannidis, 2020), and our sample size of 30 subjects per group is larger than the average neuroimaging study sample. Additionally, it would also be beneficial to have the same subjects scanned at each field strength and the imaging parameters to be as similar as possible for a more direct comparison. To our knowledge, there is limited data, if any, that has subject data scanned at 1.5, 3, and 7 T. Future studies should obtain data on the same subjects at different field strengths and apply our approach. We limited the number of included atlases due to the complexity of analyses and data processing time, but these atlases are widely used by the research community. The code on our GitHub could be used for independent replication. This tool will allow researchers to quantify the differences that are common in real data due to anatomical variation from person to person. Additionally, quantitative atlas correspondence can facilitate comparisons between studies where different atlases are used, including case-control findings on disease-specific atlases and studies of population variability.

Appendix A. Supporting information

Supplementary data associated with this article can be found in the online version at doi:10.1016/j.jneumeth.2025.110445.

Data availability

Data will be made available on request.