CerebrA: Accurate registration and manual label correction of Mindboggle-101 atlas for MNI-ICBM152 template

Accurate anatomical atlases are recognized as important tools in brain-imaging research. They are widely used to estimate disease-specific changes and therefore, are of great relevance in extracting regional information on volumetric variations in clinical cohorts in comparison to healthy populations. The use of high spatial resolution magnetic resonance imaging and the improvement in data preprocessing methods have enabled the study of structural volume changes on a wide range of disorders, particularly in neurodegenerative diseases where different brain morphometry analyses are being broadly used in an effort to improve diagnostic biomarkers. In the present dataset, we introduce the Cerebrum Atlas (CerebrA) along with the MNI-ICBM2009c average template. MNI-ICBM2009c is the most recent version of the MNI-ICBM152 brain average, providing a higher level of anatomical details. Cerebra is based on an accurate non-linear registration of cortical and subcortical labelling from Mindboggle 101 to the symmetric MNI-ICBM2009c atlas, followed by manual editing.


Background and Summary
Brain atlases are widely recognized as important tools in research for the analysis of neuroimages. High spatial resolution magnetic resonance imaging (MRI) and improved data preprocessing have enabled the study of structural volume changes on a wide range of disorders. Anatomical atlases, often averaged from multiple subjects, used to address disease-specific changes are crucial in order to provide regional information on volumetric variations in clinical cohorts in comparison to healthy populations.
The MNI-ICBM152 brain template2, from the Montreal Neurological Institute (MNI) is a crucial tool in neuroimage analysis. This multi-contrast atlas including T1w, T2w and PDw contrasts, was built recruiting brain scans from 152 young adults at 1.5 T.
The 2009 edition uses group-wise non-linear registration for better alignment of cortical structures between subjects. The MNI-ICBM152 non-linear model has many advantages. It was created from a large number of subjects; hence it represents the average anatomy of the population and is not biased unlike single-subject models. In addition, the left-right symmetric version enables interpretation of asymmetries that might be found in an analysis.
Mindboggle-101 is the largest, publicly available set of manually labelled human brain images created from 101 human scans, labelled according to a surface-based cortical labelling protocol (DKT-Desikan-Killiany-Tourville labelling protocol) 1,3. For the creation of the Mindboggle-101 dataset, developed to serve as brain atlas for use in labelling other brains, 101 T1-weighted (T1w) brain MRI images were selected and segmented based on a modification of the DKT cortical parcellation atlas3. These labels were then manually edited in agreement with the DKT protocol. Labelling was performed on the surface, yet, topographical landmarks visible in the folded surface were used to infer label boundaries. In addition, Mindboggle used non-cortical labels that were converted from Neuromorphometrics BrainCOLOR subcortex labels4.
The Cerebrum Atlas (CerebrA) includes co-registration of the Mindboggle atlas1 to the symmetric version of MNI-ICBM 2009c2 average template (at a resolution of 1 × 1 × 1 mm 3) in addition to manual editing of cortical and subcortical labels.
In the present dataset, we introduce an accurate non-linear registration of cortical and subcortical labelling from Mindboggle 101 to the symmetric MNI-ICBM2009c atlas followed by manual editing. The following preprocessing steps were applied to all MRI scans prior to building the atlas: (1) N3 non-uniformity correction 5; (2) linear normalization of each scan's intensity to the range [0-100] by a single linear histogram scaling 6; (3) automatic linear (nine parameters) registration to the ICBM 152 stereotaxic space7; and (4) brain mask creation 8. Only the voxels within the brain volume after linear mapping into stereotaxic space were used for the nonlinear registration procedure described.

MNI
The template described is generated through a hierarchical nonlinear registration procedure, with diminishing step sizes in each iteration until convergence and relies on the nonlinear registration using Automatic Nonlinear Image Matching and  Mindboggle-101. T1-w publicly accessible MRI scans were selected from 101 healthy participants. Scanner acquisition and demographic information can be found in Klein 20121 and are also available on the http://mindboggle.info/data website. The data sets that comprise the Mindboggle-101 include the 20 test-retest subjects from the "Open Access Series of Imaging Studies" data11, the 21 test-retest subjects from the "Multi-Modal ReproducibilityResource"12, with two additional subjects run under the same protocol in 3 T and 7 T scanners, 20 subjects from the "Nathan Kline Institute Test-Retest" set, 22 subjects from the "Nathan Kline Institute/Rockland Sample", the 12 "Human Language Network" subjects13, the Colin Holmes 27 template14, two identical twins, and one brain imaging colleague. T1-w MRI volumes were preprocessed and segmented and then, cortical surfaces were generated using FreeSurfer's standard recon-all image processing pipeline415,16.   19,20. We invite contributions by other researchers, in terms of alternative opinions on labeling of included structures.

Comparison between atlases. When comparing CerebrA to original labels from
Mindboggle-101 ( Figure 2) registered to ICBM152, the average Dice Kappa value was κ = 0.73 ± 0.18 (Table 1). The structures with relatively lower Dice Kappa (κ < 0.6) corresponded to the structures that needed the most correction such as the optic chiasm, inferior lateral ventricles, fourth ventricle and cerebellar vermis. The optic chiasm label was barely found in the original Mindboggle-101 registered to ICBM152 and most of it was misaligned with regards to the actual structure. To ensure that this inaccuracy was not caused by the nonlinear registration process, we further inspected the original Mindboggle-101 template and label atlas and found similar issues. For CerebrA, the optic chiasm label was redefined trying to achieve continuity amongst optic chiasma itself and optic tracts ( Figure 3, panel a). Then, the inferior lateral ventricles and fourth ventricle boundaries were improved using a threshold to differentiate CSF from parenchyma ( Figure 3, panels b and c). And finally, cerebellar vermis labels were redefined for right and left side (Figure 3, panels d-f).   Table 2.   (Figure 5b).