Test–retest reliability of brain morphology estimates
- 4.7k Downloads
Metrics of brain morphology are increasingly being used to examine inter-individual differences, making it important to evaluate the reliability of these structural measures. Here we used two open-access datasets to assess the intersession reliability of three cortical measures (thickness, gyrification, and fractal dimensionality) and two subcortical measures (volume and fractal dimensionality). Reliability was generally good, particularly with the gyrification and fractal dimensionality measures. One dataset used a sequence previously optimized for brain morphology analyses and had particularly high reliability. Examining the reliability of morphological measures is critical before the measures can be validly used to investigate inter-individual differences.
KeywordsCortical structure Subcortical Reliability Fractal dimensionality Cortical thickness Gyrification Structural complexity
A growing number of studies have investigated relationships between brain morphology and inter-individual differences. An important assumption that underlies these studies is that estimates of brain morphology are reliable. While numerous studies have investigated the test–retest reliability for estimates of cortical thickness (e.g., [1, 2, 3, 4, 5, 6, 7]) and subcortical volume (e.g., [7, 8, 9, 10, 11, 12]), the reliability of other measures of brain morphology has been less established and is an important topic of future research . Here we measured the reliability of several measures of cortical and subcortical structures; in addition to cortical thickness and subcortical volume, we examined the reliability of estimates of cortical gyrification and fractal dimensionality.
Gyrification index is a measure of the ratio between the surface area of the cortex, relative to a simulated enclosing surface that surrounds the cortex (e.g., [14, 15, 16, 17, 18]). Generally, gyrification has been suggested to be an important characteristic of the human brain [15, 16, 17, 18, 19]. In addition to the well-known differences in cortical thickness associated with age, gyrification also differs with age [20, 21, 22]; however, age-related differences in gyrification appear to have a distinct topological distribution than thickness [20, 21]. Gyrification has also been associated with a myriad of other inter-individual measures, as reviewed by Mietchen and Gaser .
Structural complexity is measured as fractal dimensionality, which uses fractal geometry principles  to measure the complexity of brain structures (see ). We recently demonstrated robust age differences in the structural complexity of cortical  and subcortical structures . Less work has been done examining the relationship between inter-individual differences and variance in complexity of cortical and subcortical regions; however, these approaches have been found to be useful in a variety of disciplines within neuroscience [25, 26].
Here we examined the test–retest reliability of several measures of brain morphology. While volumetric measures—cortical thickness and subcortical volume—have been evaluated previously, we additionally evaluated the reliability of shape-related measures, specifically gyrification and fractal dimensionality. We evaluated the FreeSurfer implementation of gyrification, as implemented by Schaer et al. . This approach generates an enclosing surface around each hemisphere and computes the ‘local’ difference in surface between this surface and the pial surface of the cortex. As such, gyrification is highest over the insula and lowest over medial cortical regions. Fractal dimensionality was evaluated based on the calcFD toolbox , which computes fractal dimensionality using intermediate files generated as part of the standard FreeSurfer pipeline. Madan and Kensinger  previously compared different algorithms for calculating fractal dimensionality using simulated 3D structures, but here we instead used multiple anatomical volumes acquired from the same participant (i.e., test–retest reliability).
Structural measurements are often used to assess longitudinal changes or inter-individual differences. For instance, advancements in measuring relationships between brain morphology and inter-individual differences have become increasingly relevant as a complementary approach to fMRI, due to aging-related confounds in group comparisons . More recently, age-related differences have been identified in BOLD signal variability [29, 30], which may be related to differences in cerebrovascular reactivity [31, 32]. As brain morphology research advances, it is critical to measure the reliability of these metrics using multiple volume acquisitions. For instance, if the effect of age on a morphological measure is small, poorer reliability may make the effect difficult to detect due to noise in the measure. A number of open-access databases include multiple scans on the same participants, enabling such reliability to be calculated. Appendix 1 summarizes a number of additional open-access datasets—in addition to those we consider here—that also include intersession test–retest reliability data.
Here we examined test–retest reliability from two open-access datasets in which participants were scanned several times over a short interval (i.e., intersession, intrascanner). In the first dataset, 30 participants were scanned 10 times within a 1-month period . In the original work, Chen et al. sought to estimate test–retest reliability of resting-state networks across intra- and inter-individual variability of six rs-fMRI measures (CCBD [Center for Cognition and Brain Disorders] dataset). In the second dataset, 69 participants were scanned twice within a 6-month period . Holmes et al. collected data for a large-scale exploration (N = 1570) of the relations among brain function, behavior, and genetics (GSP [Brain Genomics Superstruct Project] dataset). As one demonstration of the uses of this dataset, Holmes et al.  examined the relationship between cortical thickness and several measures of cognitive control.
In each of these datasets, we examined the reliability of three cortical measures: cortical thickness, gyrification, and fractal dimensionality—both of the entire cortical ribbon and across regional measures of parcellated cortex (62 regions, based on the DKT atlas; ). We additionally evaluated different approaches to calculating fractal dimensionality to establish the reliability of each of these approaches. Finally, reliability of volume and fractal dimensionality of segmented subcortical and ventricular structures also was evaluated. We consider each dataset separately, as would be the typical approach for examining test–retest reliability, and then discuss the conclusions reached using both datasets in the general discussion.
2 Study 1: CCBD
MR images were acquired using a GE MR750 3 T scanner at the Centre for Cognition and Brain Disorders (CCBD) at Hangzhou Normal University . Thirty participants aged 20–30 years old were each scanned for 10 sessions, occurring 2–3 days apart over a 1-month period. T1-weighted data were acquired using a FSPGR sequence (TR: 8.06 s; TE: 3.1 ms; flip angle: 8°; voxel size: 1.0 × 1.0 × 1.0 mm). This dataset is included as part of the Consortium for Reliability and Reproducibility (CoRR; ) as HNU1.
2.1.2 Preprocessing of the structural data
Data were analyzed using FreeSurfer 5.3.0 (https://surfer.nmr.mgh.harvard.edu) on a machine running CentOS 6.6. FreeSurfer was used to automatically volumetrically segment and parcellate cortical and subcortical structures from the T1-weighted images [37, 38, 39, 40]. FreeSurfer’s standard pipeline was used (i.e., recon-all ). No manual edits were made to the surface meshes, but surfaces were visually inspected.
Cortical thickness is calculated as the distance between the white matter surface (white–gray interface) and pial surface (gray–CSF interface) . Thickness estimates have previously been found to be in agreement with manual measurements from MRI images [41, 42], as well as ex vivo tissue measurements [43, 44]. Subcortical volume estimates have also been found to correspond well with manual segmentation protocols, particularly in young adults [45, 46, 47, 48, 49, 50, 51, 52].
Gyrification was also calculated using FreeSurfer, as described in Schaer et al. . Cortical regions were delineated based on the Desikan–Killiany–Tourville (DKT) atlas, also part of the standard FreeSurfer analysis pipeline . Intracranial volume (ICV) was also calculated using FreeSurfer .
Fractal dimensionality was quantified using the calcFD toolbox (http://cmadan.github.io/calcFD/), which we previously developed and distribute freely [21, 24]. calcFD is a MATLAB toolbox that calculates the fractal dimensionality of 3D structures and was developed to work with intermediate files from the standard FreeSurfer pipeline. Apart from when otherwise stated, FD was calculated for filled structures (FD f ) using the dilation algorithm. Here we additionally modified calcFD in two ways. First, we improved it to additionally calculate the fractal dimensionality of cortical parcellations for all regions delineated in the DKT atlas (see Appendix 2). An important consideration in decreasing the size of cortical parcellations, however, is that they inherently have decreased fractal dimensionality, i.e., becoming closer to a ‘truncated rectangular pyramid.’ Second, we adjusted the toolbox to calculate fractal dimensionality using the spherical harmonics (e.g., [54, 55, 56, 57, 58]). Additional details about this spherical harmonics approach are outlined in Appendix 3.
2.1.3 Measuring reliability
Reliability was calculated as intraclass correlation coefficient (ICC), which can be used to quantify the relationship between multiple measurements [59, 60, 61, 62]. McGraw and Wong  provide a comprehensive review of the various ICC formulas and their applicability to different research questions. ICC was calculated as the one-way random effects model for the consistency of single measurements, i.e., ICC(1). As a general guideline, ICC values between .75 and 1.00 are considered ‘excellent,’ .60–.74 is ‘good,’ .40–.59 is ‘fair,’ and below .40 is ‘poor’ . For the cortical parcellated regions, distributions of mean reliability measures (e.g., lower panel of Fig. 4) were compared using a Mann–Whitney U test, a nonparametric for testing whether two sets of values belong to the same distribution.
In the current study, we focused on regional estimates of brain morphology; a complimentary approach that we did not evaluate here is the reliability in spatial segmentation. This alternative approach evaluates the volumetric overlap between 3D structures within the same space, often quantified as a Dice coefficient (e.g., [5, 10, 48, 50, 65]). This overlap approach is often used when comparing manual and automatic segmentation protocols of the same anatomical volume; however, it can be applied to test–retest reliability by co-registering the individual anatomical volumes from the same participant to each other and comparing the resulting segmented structures’ overlap. In contrast, the present goal was to evaluate ‘summary statistics’ of the structures, such as thickness, volume, and fractal dimensionality.
2.2.1 Cortical ribbon
Test–retest reliability (ICC) for each measure and dataset, for the cortical ribbon data
Dilation filled (FD f )
Fractal dimensionality We computed the reliability of five calculations of fractal dimensionality. First, we used both the dilation and box-counting algorithms, as implemented in the calcFD toolbox, for both the filled volumes and surfaces only. We additionally used a spherical harmonics (SPHARM) approach (surface only). See Appendix 3 for further details about calculating fractal dimensionality using spherical harmonics. Figure 1 shows estimates of fractal dimensionality based on the dilation-filled approach.
As shown in Table 1, we consistently found higher reliability for the dilation algorithm than the box-counting algorithm, though this difference was not statistically significant. We found higher reliability for the spherical harmonics approach; however, this approach can only be used for surfaces of structures (rather than filled volumes).
2.2.2 Cortical parcellations
As expected (as in ), gyrification was highest in the insula and lowest over medial cortical regions (Fig. 2). Beyond this, we additionally observed greater gyrification over parietal regions, convergent with prior studies (e.g., [20, 21]). Test–retest reliability of regional gyrification was generally quite high (Figs. 3, 4) and was significantly higher for gyrification than cortical thickness [Z = 5.98, p < .001].
Regional fractal dimensionality is shown in Fig. 2. Smaller regions had lower fractal dimensionality, as smaller segmented structures inherently have less structural complexity due to both limitations MRI acquisition precision and biological constraints (also see ). Intraclass correlations (ICCs) are shown for each structural measure and brain region in Fig. 3; Fig. 4 shows the 95% confidence intervals of the inter-class correlations (ICCs) for each measure and region. Across regions, mean ICC was not significantly related to the size of the region for any of the measures [thickness: r(60) = .206, p = .11; gyrification: r(60) = .154, p = .23; fractal dimensionality: r(60) = .251, p = .05]. Test–retest reliability of regional fractal dimensionality was generally high (Figs. 3, 4) and was also significantly higher than for cortical thickness [Z = 5.46, p < .001]. Reliability did not differ between gyrification and fractal dimensionality [Z = .31, p = .75].
2.2.3 Subcortical structures
The results indicate that gyrification and fractal dimensionality have high test–retest reliability. Indeed, reliability using these measures was higher than for cortical thickness.
3 Study 2: GSP
To further assess the replicability of these findings, we calculated these same measures in a second dataset. While this dataset had only two MRI sessions, rather than 10, this dataset used an anatomical MRI sequence that was optimized for brain morphology research (based on prior validation work assessing cortical thickness and subcortical volume) [7, 68]. While this prior validation work suggests that reliability for cortical thickness and subcortical volume should be higher for this dataset, it is not clear how these improvements to volumetric measures may influence shape-related measures of morphology (i.e., gyrification and fractal dimensionality).
MR images were acquired on Siemens Trio 3 T scanners at Harvard University and Massachusetts General Hospital, as part of the Brain Genome Superstruct Project (GSP; ). This dataset includes 1570 participants from aged 18 to 25 years old. Test–retest reliability data were available for 69 participants who were scanned within 6 months of their first session (also see ). T1-weighted data were acquired using a MEMPRAGE sequence optimized for brain morphology (TR: 2.20 s; TE: 1.5, 3.4, 5.2, 7.0 ms; flip angle: 7°; voxel size: 1.2 × 1.2 × 1.2 mm) [7, 68].
3.1.2 Data analysis
The MR images were processed using an identical procedure as in Study 1. ICC was also evaluated using the same approach.
3.2.1 Cortical ribbon
3.2.2 Cortical parcellations
3.2.3 Subcortical structures
As shown in Fig. 8c, test–retest reliability was near perfect for both volume and fractal dimensionality of the subcortical structures. The regions that had relatively lower reliability (pallidum, amygdala, accumbens) were also relatively lower in Study 1, demonstrating the replicability of lower test–retest reliability in these regions—at least when segmented using FreeSurfer’s automated algorithms. Reliability was particularly high for the hippocampus and was significantly higher than in the CCBD dataset (Study 1).
Here we evaluated the test–retest reliability of several brain morphology measures using open-access datasets. Prior work had examined the reliability of volumetric measures—cortical thickness and subcortical volume; however, the present study is the first to assess reliability of shape-related measures, gyrification and fractal dimensionality.
Both datasets showed relatively high reliability for all morphology measures and additionally revealed that reliability was particularly good for the gyrification and fractal dimensionality measures. Additionally, we provide empirical evidence that the dilation approach for calculating fractal dimensionality was superior in reliability to the ‘standard’ box-counting method. These findings held across two datasets, but reliability was particularly good in the GSP dataset, where the anatomical sequence had been previously optimized for use in brain morphology studies.
Although reliability was good in these datasets, there is still the question of how reliability may be increased in future studies. A number of factors have been found to influence estimates of brain morphology. Broadly, these factors can be divided into three categories: MR acquisition, biological, and analysis related. For MR acquisition, there are not yet enough datasets available to systematically examine how reliability is affected by the particular acquisition protocols, although the current data suggest that sequences previously optimized for brain morphology analyses (i.e., those used in GSP dataset) will have better reliability. Another acquisition-related factor is head movement; movement has been shown to lead to decreased estimates of cortical thickness [69, 70, 71, 72], though it is unclear how movement would affect measures of gyrification and fractal dimensionality. This issue may become less critical in future studies, as recent advances in structural imaging have been able to attenuate movement-related artifacts (e.g., [73, 74, 75, 76]). Morphological measures can also be influenced by biological confounds, such as hydration [77, 78, 79, 80] or circadian rhythms [81, 82]. Additionally, it is important to control for variations in analysis software and operating system, which can also affect brain morphology estimates [65, 83, 84].
While the surface reconstructions were visually inspected, the surfaces were not manually edited, for two reasons. First and foremost, the quality of the automatic reconstructions was judged to be acceptable and did not require manual intervention. While manual editing is more necessary with older adult and patient populations, all of the individuals included in the present work were young adults. Additionally, manual editing introduces a subjective component and is often not conducted in studies of reconstruction reliability [2, 5, 6, 46], though some reliability studies have included minimal manual editing [4, 7]. Given that no manual editing was conducted, the reliability estimates presented here may serve as a lower bound, where manual editing would be expected to increase reliability [4, 6]; however, there is evidence that editing may not sufficiently influence regional estimates [85, 86].
Fractal dimensionality was used here as a measure of the complexity in the shape of a structure. Results indicate that this measure was generally more reliable than volumetric morphological measures, likely because fractal dimensionality is influenced by both shape and volumetric characteristics that often covary [21, 24, 87, 88, 89]. By pooling from both of these characteristics, fractal dimensionality appears to be more reliable and should be considered in future research investigating the relationship between brain morphology and inter-individual differences.
In sum, here we evaluated the reliability of several brain morphology estimates using two open-access datasets. Reliability was generally high, providing support for using gyrification and fractal dimensionality measures to evaluate inter-individual or between-sample differences in morphology.
Portions of this research were supported by a grant from the National Institutes of Health (MH080833; to E.A.K.) and by funding provided by Boston College. C.R.M. was supported by a fellowship from the Canadian Institutes of Health Research (FRN-146793). MRI data used in the preparation of this article were obtained from several sources, data were provided in part by: (1) the Center for Cognition and Brain Disorders (CCBD; ) as dataset HNU1 in the Consortium for Reliability and Reproducibility (CoRR; ); and (2) the Brain Genomics Superstruct Project (GSP; ) of Harvard University and the Massachusetts General Hospital (Principal Investigators: Randy Buckner, Joshua Roffman, and Jordan Smoller), with support from the Center for Brain Science Neuroinformatics Research Group, the Athinoula A. Martinos Center for Biomedical Imaging, and the Center for Human Genetic Research, and 20 individual investigators at Harvard and MGH generously contributed data to the overall project.
Compliance with ethical standards
Conflict of interest
On behalf of all authors, the corresponding author states that there is no conflict of interest.
- 1.Dickerson BC, Fenstermacher E, Salat DH, Wolk DA, Maguire RP, Desikan R et al (2008) Detection of cortical thickness correlates of cognitive performance: reliability across MRI scan sessions, scanners, and field strengths. NeuroImage 39:10–18. doi: 10.1016/j.neuroimage.2007.08.042 CrossRefGoogle Scholar
- 2.Han X, Jovicich J, Salat D, van der Kouwe A, Quinn B, Czanner S et al (2006) Reliability of MRI-derived measurements of human cerebral cortical thickness: the effects of field strength, scanner upgrade and manufacturer. NeuroImage 32:180–194. doi: 10.1016/j.neuroimage.2006.02.051 CrossRefGoogle Scholar
- 7.Wonderlick JS, Ziegler DA, Hosseini-Varnamkhasti P, Locascio J, Bakkour J, van der Kouwe A et al (2009) Reliability of MRI-derived cortical and subcortical morphometric measures: effects of pulse sequence, voxel geometry, and parallel imaging. NeuroImage 44:1324–1333. doi: 10.1016/j.neuroimage.2008.10.037 CrossRefGoogle Scholar
- 10.Jovicich J, Czanner S, Han X, Salat D, van der Kouwe A, Quinn B et al (2009) MRI-derived measurements of human subcortical, ventricular and intracranial brain volumes: reliability effects of scan sessions, acquisition sequences, data analyses, scanner upgrade, scanner vendors and field strengths. NeuroImage 46:177–192. doi: 10.1016/j.neuroimage.2009.02.010 CrossRefGoogle Scholar
- 17.Zilles K, Armstrong E, Schleicher A, Kretschmann H-J (1988) The human pattern of gyrification in the cerebral cortex. Anat Embryol 179:173–179. doi: 10.1007/BF00304699
- 18.Zilles K, Armstrong E, Moser KH, Schleicher A, Stephan H (1989) Gyrification in the cerebral cortex of primates. Brain Behav Evol 34:143–150. doi: 10.1159/000116500
- 22.Magnotta VA, Andreasen NC, Schultz SK, Harris G, Cizadlo T, Heckel D et al (1999) Quantitative in vivo measurement of gyrification in the human brain: changes associated with aging. Cereb Cortex 9:151–160. doi: 10.1093/cercor/9.2.151
- 49.Mulder ER, de Jong RA, Knol DL, van Schijndel RA, Cover KS, Visser PJ et al (2014) Hippocampal volume change measurement: quantitative assessment of the reproducibility of expert manual outlining and the automated methods FreeSurfer and FIRST. NeuroImage 92:169–181. doi: 10.1016/j.neuroimage.2014.01.058 CrossRefGoogle Scholar
- 53.Buckner RL, Head D, Parker J, Fotenos AF, Marcus D, Morris JC, Snyder AZ (2004) A unified approach for morphometric and functional data analysis in young, old, and demented adults using automated atlas-based head size normalization: reliability and validation against manual measurement of total intracranial volume. NeuroImage 23:724–738. doi: 10.1016/j.neuroimage.2004.06.018 CrossRefGoogle Scholar
- 54.Chung MK (2014) Statistical and computational methods in brain image analysis. CRC Press, New YorkGoogle Scholar
- 59.Asendorpf J, Wallbott HG (1979) Maße der Beobachterübereinstimmung: ein systematischer Vergleich. Zeitschrift für Sozialpsychologie 10:243–252Google Scholar
- 76.Tisdall MD, Reuter M, Qureshi A, Buckner RL, Fischl B, van der Kouwe AJW (2016) Prospective motion correction with volumetric navigators (vNavs) reduces the bias and variance in brain morphometry induced by subject motion. NeuroImage 127:11–22. doi: 10.1016/j.neuroimage.2015.11.054 CrossRefGoogle Scholar
- 79.Nakamura K, Brown RA, Araujo D, Narayanan S, Arnold DL (2014) Correlation between brain volume change and T2 relaxation time induced by dehydration and rehydration: implications for monitoring atrophy in clinical studies. NeuroImage Clin 6:166–170. doi: 10.1016/j.nicl.2014.08.014 CrossRefGoogle Scholar
- 84.Gronenschild EHBM, Habets P, Jacobs HIL, Mengelers R, Rozendaal N, van Os J, Marcelis M (2012) The effects of FreeSurfer version, workstation type, and Macintosh operating system version on anatomical volume and cortical thickness measurements. PLoS ONE 7:e38234. doi: 10.1371/journal.pone.0038234 CrossRefGoogle Scholar
- 87.Gerig G, Styner M, Jones D, Weinberger D, Lieberman J (2001b) Shape analysis of brain ventricles using SPHARM. In: Proceedings of the IEEE workshop on mathematical methods in biomedical image analysis (MMBIA 2001), pp 171–178. doi: 10.1109/mmbia.2001.991731
- 97.Styner M, Oguz I, Xu S, Brechbühler C, Pantazis D, Levitt JJ et al (2006) Framework for the statistical shape analysis of brain structures using SPHARM-PDM. Insight Journal 1071:242–250Google Scholar
- 98.Yu P, Yeo BTT, Grant PE, Fischl B, Golland P (2007) Cortical folding development study based on over-complete spherical wavelets. In: International Conference on Computer Vision 2007 proceedings of the workshop on mathematical methods in biomedical image analysis (MMBIA). doi: 10.1109/iccv.2007.4409137
- 108.Froeling M, Tax CMW, Vos SB, Luijten PR, Leemans A (in press) “MASSIVE” brain dataset: multiple acquisitions for standardization of structural imaging validation and evaluation. Magn Reson Med. doi: 10.1002/mrm.26259
Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.