Skip main navigation
Close Drawer MenuOpen Drawer Menu
Search
Quick Search in Journals
Quick Search anywhere
Advanced Search
Menu
  • Information
    • Previous
    Next
    Original ResearchFree Access

    Leveraging Clinical Imaging Archives for Radiomics: Reliability of Automated Methods for Brain Volume Measurement

    Published Online:Apr 27 2017https://doi.org/10.1148/radiol.2017161928

    Abstract

    Purpose

    To validate the use of thick-section clinically acquired magnetic resonance (MR) imaging data for estimating total brain volume (TBV), gray matter (GM) volume (GMV), and white matter (WM) volume (WMV) by using three widely used automated toolboxes: SPM (www.fil.ion.ucl.ac.uk/spm/), FreeSurfer (surfer.nmr.mgh.harvard.edu), and FSL (FMRIB software library; Oxford Centre for Functional MR Imaging of the Brain, Oxford, England, https://fsl.fmrib.ox.ac.uk/fsl).

    Materials and Methods

    MR images from a clinical archive were used and data were deidentified. The three methods were applied to estimate brain volumes from thin-section research-quality brain MR images and routine thick-section clinical MR images acquired from the same 38 patients (age range, 1–71 years; mean age, 22 years; 11 women). By using these automated methods, TBV, GMV, and WMV were estimated. Thin- versus thick-section volume comparisons were made for each method by using intraclass correlation coefficients (ICCs).

    Results

    SPM exhibited excellent ICCs (0.97, 0.85, and 0.83 for TBV, GMV, and WMV, respectively). FSL exhibited ICCs of 0.69, 0.51, and 0.60 for TBV, GMV, and WMV, respectively, but they were lower than with SPM. FreeSurfer exhibited excellent ICC of 0.63 only for TBV. Application of SPM’s voxel-based morphometry on the modulated images of thin-section images and interpolated thick-section images showed fair to excellent ICCs (0.37–0.98) for the majority of brain regions (88.47% [306924 of 346916 voxels] of WM and 80.35% [377 282 of 469 502 voxels] of GM).

    Conclusion

    Thick-section clinical-quality MR images can be reliably used for computing quantitative brain metrics such as TBV, GMV, and WMV by using SPM.

    © RSNA, 2017

    Online supplemental material is available for this article.

    Introduction

    Current radiology practice primarily relies on treating images as pictures, and interpretations are made on the basis of visual inspection of the images (1). According to Gillies et al, radiomics refers to the “extraction of quantitative features that result in the conversion of images into mineable data and the subsequent analysis of these data for decision support” (1). Brain volume estimation, referred to as brain volumetry, is often used in radiomics to estimate brain tissue volumes. Automated brain volumetry is increasingly used on structural magnetic resonance (MR) images in both research and clinical domains to help diagnose disease, track disease progression, and monitor effects of treatment (2). In the clinical setting, routine head MR images are commonly acquired with lower spatial resolution and higher section thicknesses to enable faster acquisition times. It is unclear, however, if low-resolution clinical brain MR images are limited in their use with automated segmentation tools that have traditionally been used with high-resolution images (ie, usually with a section thickness of <2 mm). Specifically, we do not know how brain volume metrics derived from thick-section images compare with those derived from thin-section images. This is important because establishing the reliability of clinical MR imaging data for research-driven volumetric analyses will allow for the use of vast archives of previously unused clinical images.

    Brain volumetry typically involves segmentation and quantification of various tissue types of the brain such as gray matter (GM), white matter (WM), and cerebrospinal fluid. Other metrics include total intracranial volume and total brain volume (TBV). Automated segmentation tools are widely applied because of the labor-intensive nature of manual segmentation. A variety of automated methods currently exist, of which SPM (www.fil.ion.ucl.ac.uk/spm/) (3), FreeSurfer (surfer.nmr.mgh.harvard.edu) (4), and FSL (FMRIB software library; Oxford Centre for Functional MR Imaging of the Brain, Oxford, England, https://fsl.fmrib.ox.ac.uk/fsl) (5) are among the most popular. A study by Helms (6) provided an extensive review of these and other available brain segmentation methods.

    To date, few studies performed brain volumetry by using thick-section images. Smith et al (5,7) used FSL on MR images of varying section thicknesses (1–6 mm) from the same patients and found that FSL estimates of TBV did not vary with section thickness. Eritaia et al (8) examined the effect of sparse sampling of image sections and showed that reliable estimates of total intracranial volume can be achieved up to a sampling density of one in 25 sections. These results were confirmed in a recent study (9). Klauschen et al (10) compared the performances of SPM version 5, FSL, and FreeSurfer in calculating GM volume (GMV), WM volume (WMV), and TBV by using thin-section images. This study found that volumetric accuracy of SPM version 5 and FSL was better than that of FreeSurfer. A more recent study (11) showed that SPM version 12 performed better than FreeSurfer for calculation of total intracranial volume. However, the reliability of applying automated methods on clinical-quality MR images has not been well established in the literature. In this study, we aim to validate the use of thick-section clinically acquired MR images for estimating GMV, WMV, and TBV by using three widely used automated methods (SPM, FreeSurfer, and FSL).

    Materials and Methods

    Patients

    This study was reviewed and approved by Geisinger Health System Foundation institutional review board. The data used in this study were not identifiable and no protected health information was collected, accessed, used, or distributed. This study was part of a larger research initiative that addressed the question of leveraging clinical imaging archives for research studies. As part of that initiative, we deidentified 2500 randomly selected patients with head MR images from our clinical picture archiving and communication system; all images were acquired between March and November of 2014. Of these patients, 44 had both thick- and thin-section images with complete head coverage acquired from the same imager in the same imaging session. Of the 44 patients, 38 were free of intracranial abnormalities on the basis of a neuroradiologist’s clinical review (G.J.M., with 6 years of experience). The MR images of these 38 patients (mean age, 22 years; age range, 1–71 years; 11 female patients) were used as the final dataset of this study. A retrospective inspection of the deidentified reports indicated that these 38 images were acquired for clinical purposes as part of our institution’s routine clinical imaging protocol for evaluating patients with seizures or reported headaches.

    Image Acquisition

    Twenty-two of the patients were imaged by using an Achieva 1.5-T imager (Philips, Best, the Netherlands) and 16 patients were imaged by using a Signa HDxt 1.5-T imager (GE Medical Systems, Milwaukee, Wis). Further information on image parameters is available in Table 1.

    Table 1 Imager Parameters

    Table 1

    Note.—SPGR = spoiled gradient recalled, SE = spin echo.

    Brain Volume Calculations

    To estimate brain volumes by using SPM, we applied the unified segmentation algorithm (3), provided as the Segment tool in SPM (version 12). GMV and WMV were calculated according to Approach 2 outlined in Malone et al (11) by using the native space probabilistic tissue maps produced during segmentation. FreeSurfer volumes were obtained by using the recon-all–all pipeline of FreeSurfer (version 6 beta). The total GMV and cerebral WMV values found in the aseg.stats output file of FreeSurfer were used for GMV and WMV, respectively. Further information on the FreeSurfer segmentation process can be found in Fischl et al (12). In FSL (version 5.0.8) segmentation, volumes were obtained by using SIENAX (5). We used the nonnormalized GMV and WMV volumes produced by SIENAX. All images were processed by using the default parameters of the toolboxes.

    Voxel-based Morphometry

    Voxel-based morphometry in SPM uses the modulated GM and WM maps produced by the unified segmentation pipeline (13). Modulated images allow users to compare regional tissue density in a standard space and absolute volume differences across patients (14). To examine if thick-section clinical images can be reliably used for voxel-based morphometry, we obtained modulated images for GM and WM for both thick- and thin-section MR images by using the modulated option for warped tissue selection in the segment tool in SPM. Modulated images produced from thick-section images had aliasing artifacts because of low resolution. To remove this artifact, we resectioned the thick-section images to 1 mm by using nearest-neighbor interpolation before running the unified segmentation pipeline.

    Statistical Methods

    GMV, WMV, and TBV (GMV + WMV) were obtained for thick- and thin-section images by using SPM, FreeSurfer, and FSL. The reliability and level of agreement between thick- and thin-section image volumes were evaluated by using intraclass correlation coefficient (ICC) for all three volumes. ICC was computed by using one-way random effects model (case-1, as defined in McGraw et al [15]) with patients (ie, row effects) as random effects assuming a normal distribution. Before applying ICC, volumes estimated by the automated methods were tested for normal distribution by using Kolmogorov-Smirnov test (16). Reliability was classified on the basis of ICCs with the following scale: poor, 0–0.36; fair, 0.37–0.47; good, 0.48–0.55; and excellent, 0.56–1.0. For our sample size of 38 images, fair, good, and excellent corresponded to P values of less than .01, less than .001, and less than .0001, respectively. The difference in volumes between thick- and thin-section images was compared by using the percentage difference and Bland-Altman plots (17). Percentage differences between the thick- and thin-section estimates were calculated as a percentage of the thin-section estimate. Figure 1 outlines the intermethod comparison methodologic analysis.

    Figure 1:

    Figure 1: Flowchart of the intermethod brain volume comparison between thick- and thin-section MR images. Green arrows represent the raw image input to three different automated volume estimation methods. Orange arrows represent estimated brain volumes. The volume comparison box represents performance of statistical analyses to compare thick- and thin-section image volumes. The intermethod comparison box (gray box) represents the comparison of performance between the three methods.

    Download as PowerPointOpen in Image Viewer

    In addition to the reliability tests and after establishing the most reliable method, analyses were performed by using estimates from the most reliable method to study the effect of age and imager heterogeneity. The effect of age on GM and WM tissue intensity contrast in structural MR images was previously demonstrated by several studies (18,19). To evaluate the effect of age on the reliability of thick-section brain volume estimates, the following experiment was performed: ICCs between thick- and thin-section volume estimates were iteratively calculated starting with the 10 youngest patients (age, 1–4 years) and then sequentially adding the next-older patient to the group.

    The structure and signal intensities of infant MR images are different from those of adults. To verify whether the infants influenced the reliability, ICCs were calculated between thick- and thin-section volumes for the cohort after excluding the infants (age, <2 years; n = 5).

    MR image signal intensity range, noise, and tissue contrast are different for images acquired by using different imagers. The effect of imager heterogeneity on reliability is verified by computing ICCs for each imager separately (Achieva, n = 22; Signa, n = 16).

    For voxel-based morphometry analysis, voxel-by-voxel ICC was performed between the voxels of the modulated images obtained from thick- and thin-section images for GM and WM maps. The voxel-by-voxel analysis created a stereotaxic map of voxels where the tissue concentrations were reliably reproduced. As in previous studies (20,21), voxels with 10% or greater probability of belonging to GM or WM were selected for analysis. The tissue probability was provided by the tissue probability map distributed with the SPM package. The tissue probability map is defined in the Montreal Neurologic Institute (22) space. All statistical analyses were performed by using statistical software (Matlab version 8.6.0; Mathworks, Natick, Mass).

    Results

    Brain Volumetry

    Among the 38 patients included in the study, FreeSurfer processing failed for three patients. FreeSurfer failed to automatically register images of two patients to the atlas. The processing of the third patient had to be manually terminated because the processing time exceeded 36 hours. These three patients were excluded from the FreeSurfer analysis.

    The performance of the three automated methods between thick-section and thin-section image volumes is presented in Table 2. All volumes estimated by the methods satisfied the criterion for normal distribution as determined by the Kolmogorov-Smirnov test. SPM showed excellent reliability between thick- and thin-sectioned image volumes for TBV, GMV, and WMV (ICCs were 0.97, 0.85, and 0.83, respectively). FSL exhibited excellent reliability for TBV and WMV (ICCs were 0.69 and 0.60, respectively) and good reliability for GMV (ICC, 0.51), but ICC values were lower than those of SPM. FreeSurfer showed the lowest reliability among the methods for all the volumes and had excellent reliability only for TBV (ICC, 0.63) and poor reliability for GMV and WMV (ICCs were 0.30 and 0.16, respectively). Average GMV estimates indicated that SPM was the largest (0.70 L) and FreeSurfer was the lowest (0.52 L). In WMV, however, SPM had the lowest estimate for WMV (0.38 L), whereas FSL exhibited the highest (0.56 L). One outlier (Fig 2) was observed in FSL, which exhibited a thick-section GMV of 1.423 L, which was more than 4 standard deviations away from the mean thick-section GMV of FSL (0.63 L). The GMVs from the same patient’s thick-section image for SPM and FreeSurfer were 0.58 L and 0.43 L, respectively, which were within 1 standard deviation from their respective means. Figure 2 illustrates volumes derived from thick-section images plotted against thin-section images for TBV, GMV, and WMV for all three methods with trend lines and reference lines. Points with identical estimates from thick-section and thin-section volumes should fall at the reference line. SPM showed the best linear trend among the three methods for all the volumes followed by FSL. FreeSurfer showed large deviations from the trend line for all the volumes. We also compared the three methods after excluding from SPM and FSL the three patients for whom FreeSurfer failed, and we observed similar results (Table E1 [online]).

    Table 2 Summary of Statistical Analysis on Thick- and Thin-Section Volume Estimations

    Table 2

    Note.—Data are mean ± standard deviation. There were 38 patients for SPM, 35 patients for FreeSurfer, and 38 patients for FSL.

    Figure 2:

    Figure 2: Scatterplots between thick-section (y-axis) and thin-section estimates (x-axis) of TBV, GMV, and WMV as estimated by three different automated methods. Blue lines represent the trend lines fitted to the scatter points. Black lines represent the y = x reference lines.

    Download as PowerPointOpen in Image Viewer

    SPM exhibited the lowest mean and standard deviation of the percentage difference for all three volumes (Table 2). The difference between thick-section and thin-section volumes was plotted against their average in Bland-Altman plots (Figure 3). SPM showed the lowest standard deviation of the volume error (thick – thin volume) for the three brain volumes. The trend line fitted between volume difference versus mean volume of the thick- and thin-section pair showed no particular pattern for all three methods. The mean of the difference between thick- and thin-section volumes represents the bias introduced by the methods.

    Figure 3:

    Figure 3: Bland-Altman plots showing thick minus thin volume difference (y-axis) plotted against the respective mean value (x-axis) of thick and thin volumes for each patient for TBV, GMV, and WMV estimated by three automated volume estimation. Solid blue line represents the trend lines. Numerical values of the mean difference (red line) and ±2 standard deviations (dashed blue line) are also shown.

    Download as PowerPointOpen in Image Viewer

    The effect of age on SPM reliability is provided in Figure E1 (online). Results indicate that the reliability of TBV is stable with increasing age, but the reliability of GMV and WMV declined marginally with increasing age. However, the reliability for all three volumes remained excellent at all ages. When ICCs were calculated for SPM estimates after excluding the infants (age, <2 years; n = 5) from the cohort, the volumes showed excellent agreement. After the removal of infants, the ICC did not change for TBV (ICC, 0.97), ICC marginally improved from 0.85 to 0.86 for GMV, and ICC marginally declined from 0.83 to 0.78 for WMV.

    When the effect of imager heterogeneity was assessed by using SPM estimates, the reliability was robust for each imager separately despite lower sample size. For the Philips imager (n = 22) reliability was excellent for TBV, GMV, and WMV (ICC was 0.99, 0.90, and 0.85, respectively; P < .0001). The reliability of the GE imager (n = 16) was excellent for TBV (ICC, 0.90; P < .0001), but reliability decreased but remained statistically significant for WMV (ICC, 0.74; P < .001) and GMV (ICC, 0.57; P < .01).

    Voxel-based Morphometry

    ICC for GM and WM tissue regions between thick- and thin-section images are presented in Figure 4. The sagittal and axial maps of the GM (Fig 4a) and WM (Fig 4b) are shown; red regions represent significantly correlated voxels (n = 38; P < .01; fair to excellent reliability; ICC, >0.37) and green regions show voxels with ICC of 0.36 or less. Of the GM and WM voxels, 80.35% (377 282 of 469 502 voxels; ICC, 0.37–0.97) and 88.47% (306 924 of 346 916 voxels; ICC, 0.37–0.98), respectively, show agreement between thick- and thin-section images. GM voxels above the spinal cord and GM–WM boundary in the cerebellum and near the ventricles showed poor reliability. In WM, mismatches were observed in voxels near the ventricles and GM-WM boundary in the cerebellum.

    Figure 4:

    Figure 4: Sagittal and axial views of ICC between thick- and thin-section MR images in SPM voxel-based morphometry for, A, GM and, B, WM. All red regions represent fair to excellent agreement (ICC, >0.37; P < .01; n = 38) and all green regions represent insignificant ICC. The section coordinates are in Montreal Neurologic Institute space.

    Download as PowerPointOpen in Image Viewer

    Discussion

    In this study we compared brain volume estimates from thick-section (>3 mm) MR images to those obtained from thin-section (<2 mm) MR images from routine clinical imaging. We evaluated the reliability of three separate automated toolboxes (SPM, FSL, and FreeSurfer) in estimating TBV, GMV, and WMV by using thick-section brain images. The best method is characterized by excellent reliability and lowest bias. With this criterion, SPM was the superior overall performer for all three volumes examined. FSL exhibited good to excellent agreement, albeit much lower than SPM. However, FreeSurfer exhibited excellent agreement for TBV only and poor agreement for GMV and WMV. In voxel-based morphometry with SPM, our results indicated fair to excellent reliability between thick- and thin-section images for GM and WM maps for the majority of brain voxels. Our findings suggest that thick-section MR images can be reliably used for estimating TBV and brain tissue volumes by using SPM, a widely applied automated method. Additionally, clinical quality images may be used for voxel-based morphometry in SPM.

    Previous work has not directly attempted to compare thick- and thin-section MR image reliability for computing brain volumes, although several studies (8,9) examined the effect of sparse sampling of MR image sections for manually computing total intracranial volume. These studies indicated that sparse sampling can provide accurate total intracranial volume estimates. Similarly, our findings indicate that thick-section images can provide reliable brain volume estimates by using automated methods. Previous studies (10,11) indicated that SPM gives the most accurate estimates for brain volumes compared with volumes calculated from manual segmentation. Our study built on these findings, and we report that SPM gives the most reliable estimates for TBV, GMV, and WMV among the three methods when thick-section images were used.

    In addition, we demonstrated that the reliability of automated methods for thin-versus-thick estimation for TBV was stable with increasing age, but declined marginally for GMV and WMV. Although there was a marginal decline of ICC for GMV and WMV, all ICC values remained excellent. The decline in GMV and WMV reliability can be attributed to decreasing GM and WM tissue intensity contrast with age (18,19).

    There are two important limitations to address regarding our findings. First, images were acquired from different imagers and had varying acquisition planes and acquisition sequences. In spite of this heterogeneity, consistent results emerged. Therefore, we do not expect that differences in imagers and imager parameters influenced the results. The second limitation is that our study has a small sample size and our findings need to be further validated with a large dataset.

    Despite the presence of large archives of clinical-quality MR images, to our knowledge this source of valuable data has not been used to its full extent in research studies. Our study demonstrates the potential to use large existing clinical brain imaging archives for radiomics, which opens a vast new data resource to researchers and clinicians for a variety of studies. These include clinical outcome and association studies, studies of pharmacologic efficacy, and phenome- and genome-wide association studies.

    Advances in Knowledge

    • ■ Automated tools applied to thick-section MR images reliably estimated brain volumes compared with thin-section images from the same patients with excellent intraclass correlation coefficient (ICC) (ICC, 0.83–0.97).

    • ■ Of the three automated toolboxes evaluated, the statistical parametric mapping (SPM; www.fil.ion.ucl.ac.uk/spm/) toolbox performed most reliably for all three brain volumes; total brain volume (ICC, 0.97), gray matter volume (ICC, 0.85), and white matter volume (ICC, 0.83).

    • ■ In voxel-based morphometry of SPM, tissue density for the majority of brain voxels (>80%) could also be reliably estimated by using thick-section MR images.

    Implication for Patient Care

    • ■ Automated methods can be used for brain volume estimations on standard-of-care clinical-quality brain MR images.

    Disclosures of Conflicts of Interest: V.R.A. disclosed no relevant relationships. A.M.M. disclosed no relevant relationships. M.H. disclosed no relevant relationships. S.A.B. disclosed no relevant relationships. G.J.M. disclosed no relevant relationships.

    Author Contributions

    Author contributions: Guarantors of integrity of entire study, V.R.A., A.M.M.; study concepts/study design or data acquisition or data analysis/interpretation, all authors; manuscript drafting or manuscript revision for important intellectual content, all authors; approval of final version of submitted manuscript, all authors; agrees to ensure any questions related to the work are appropriately resolved, all authors; literature research, V.R.A., M.H., S.A.B., G.J.M.; experimental studies, V.R.A., A.M.M.; statistical analysis, V.R.A., A.M.M., M.H., S.A.B.; and manuscript editing, all authors

    References

    • 1. Gillies RJ, Kinahan PE, Hricak H. Radiomics: images are more than pictures, they are data. Radiology 2016;278(2):563–577. Link, Google Scholar
    • 2. Giorgio A, De Stefano N. Clinical use of brain volumetry. J Magn Reson Imaging 2013;37(1):1–14. Crossref, Medline, Google Scholar
    • 3. Ashburner J, Friston KJ. Unified segmentation. Neuroimage 2005;26(3):839–851. Crossref, Medline, Google Scholar
    • 4. Dale AM, Fischl B, Sereno MI. Cortical surface-based analysis. I. Segmentation and surface reconstruction. Neuroimage 1999;9(2):179–194. Crossref, Medline, Google Scholar
    • 5. Smith SM, Zhang Y, Jenkinson M, et al. Accurate, robust, and automated longitudinal and cross-sectional brain change analysis. Neuroimage 2002;17(1):479–489. Crossref, Medline, Google Scholar
    • 6. Helms G. Segmentation of human brain using structural MRI. MAGMA 2016;29(2):111–124. Crossref, Medline, Google Scholar
    • 7. Smith SM. Fast robust automated brain extraction. Hum Brain Mapp 2002;17(3):143–155. Crossref, Medline, Google Scholar
    • 8. Eritaia J, Wood SJ, Stuart GW, et al. An optimized method for estimating intracranial volume from magnetic resonance images. Magn Reson Med 2000;44(6):973–977. Crossref, Medline, Google Scholar
    • 9. Sargolzaei S, Sargolzaei A, Cabrerizo M, et al. A practical guideline for intracranial volume estimation in patients with Alzheimer’s disease. BMC Bioinformatics 2015;16(Suppl 7):S8. Crossref, Medline, Google Scholar
    • 10. Klauschen F, Goldman A, Barra V, Meyer-Lindenberg A, Lundervold A. Evaluation of automated brain MR image segmentation and volumetry methods. Hum Brain Mapp 2009;30(4):1310–1327. Crossref, Medline, Google Scholar
    • 11. Malone IB, Leung KK, Clegg S, et al. Accurate automatic estimation of total intracranial volume: a nuisance variable with less nuisance. Neuroimage 2015;104:366–372. Crossref, Medline, Google Scholar
    • 12. Fischl B, Salat DH, Busa E, et al. Whole brain segmentation: automated labeling of neuroanatomical structures in the human brain. Neuron 2002;33(3):341–355. Crossref, Medline, Google Scholar
    • 13. Mechelli A, Price CJ, Friston KJ, Ashburner J. Voxel-based morphometry of the human brain: Methods and applications. Curr Med Imaging Rev 2005;1(2):105–113. Crossref, Google Scholar
    • 14. Good CD, Johnsrude IS, Ashburner J, Henson RN, Friston KJ, Frackowiak RS. A voxel-based morphometric study of ageing in 465 normal adult human brains. Neuroimage 2001;14(1 Pt 1):21–36. Crossref, Medline, Google Scholar
    • 15. McGraw KO, Wong SP. Forming inferences about some intraclass correlation coefficients. Psychol Methods 1996;1(1):30. Crossref, Google Scholar
    • 16. Massey FJ. The Kolmogorov-Smirnov Test for Goodness of Fit. J Am Stat Assoc 1951;46(253):68–78. Crossref, Google Scholar
    • 17. Bland JM, Altman DG. Statistical methods for assessing agreement between two methods of clinical measurement. Lancet 1986;1(8476):307–310. Crossref, Medline, Google Scholar
    • 18. Kim DM, Xanthakos SA, Tupler LA, et al. MR signal intensity of gray matter/white matter contrast and intracranial fat: effects of age and sex. Psychiatry Res 2002;114(3):149–161. Crossref, Medline, Google Scholar
    • 19. Knight MJ, McCann B, Tsivos D, Couthard E, Kauppinen RA. Quantitative T1 and T2 MRI signal characteristics in the human brain: different patterns of MR contrasts in normal ageing. MAGMA 2016;29(6):833–842. Crossref, Medline, Google Scholar
    • 20. Peelle JE, Cusack R, Henson RNA. Adjusting for global effects in voxel-based morphometry: gray matter decline in normal aging. Neuroimage 2012;60(2):1503–1516. Crossref, Medline, Google Scholar
    • 21. Lorio S, Lutti A, Kherif F, et al. Disentangling in vivo the effects of iron content and atrophy on the ageing human brain. Neuroimage 2014;103:280–289. Crossref, Medline, Google Scholar
    • 22. Tzourio-Mazoyer N, Landeau B, Papathanassiou D, et al. Automated anatomical labeling of activations in SPM using a macroscopic anatomical parcellation of the MNI MRI single-subject brain. Neuroimage 2002;15(1):273–289. Crossref, Medline, Google Scholar

    Article History

    Received September 8, 2016; revision requested November 21; revision received December 23; accepted January 13, 2017; final version accepted February 9.
    Published online: Apr 27 2017
    Published in print: Sept 2017