Average models of cortical structure are widely used to both analyse and report findings, they are synonymous with neuroimaging studies of functional and structural anatomy. Probably the first model of average was built by Alan Evans in 1992, it was an average model of 250 young normals. Given average models widespread use, it is no surprise that many complementary and indeed often competing models have surfaced in the ensuing years. The MNI itself has released three further models since the original 205: the average 305, colin 27 and ICBM152. Some of these models have also included extra data beyond the intensity averages such as segmentation probability maps and manual tracings of sub-structures. This page is an attempt to describe in detail the background and motivation for each of these and provide guidance concerning the choice and use of these various models for the neuroimaging community.
Originally, average models of anatomy in paper or digital form were most commonly used for stereotactic surgical procedures. Since then a more prevalent use has emerged, their utility as a means to compare results between studies, groups and centres has proved invaluable to the neuroimaging community. In this case, instead of matching an atlas to the individual (stereotaxy), novel data is matched to a model thus providing a common co-ordinate system in which comparisons can be made and inferences regarding average morphology can be made. Automatic methods have been developed to further the latter goal with the advantage of removing operator bias inherent with manual point based registration.
Models also allow us to perform non-subjective analysis of subregions within the cortex. This is made possible by the average structure that is present within the model. Obviously, there are differences between individuals and thus differing models have been developed with differing intended purposes. Some of the most commonly used models and their intended purposes are outlined below.
Talairach and Tournoux spaceEdit
The most commonly reported model of cortical anatomy within the neuroimaging community is the so-called Talairach and Tournoux space. The original Talairach and Tournoux atlas was a printed paper atlas developed for stereotactic procedures (e.g.: Pallidotomy and Thallidotomies) and thus exhibits good accuracy in the central region of the cortex although this is often at the expense of accuracy in others areas of the cortex, most particularly on the cortical surface. This comes as no surprise given that the AC-PC line alignment technique was developed specifically to align the central region of the cortex as part of surgical planning.
Talairach and Tournoux developed their atlas in the absence of cytoarchitectonic information and generated it from two series of sections from a single 60 year old female brain. One half was sectioned sagitally and the other coronally. The transverse images in the atlas were manually approximated from the information obtained in the sagittal and coronal planes. As the majority of normal brains exhibit some left-right asymmetry there are many areas within the atlas itself that are not self-consistent between the three views, e.g.: area 44 and area 9. For this reason, the original atlas also makes no differentiation between the left and right hemispheres of the cortex. Given the original goals of the atlas (Pallidotomy and Thallidotomies), this was and remains more than sufficient for this purpose.
A strict mapping to Talairach space involves manual identification of the AC-PC line and associated landmarks and then application of three piecewise linear scaling factors (). This technique is difficult to apply and often leads to error in localisation due to residual anatomical variability in the co-ordinate system (Drevits et al 1988). The earliest published use of Talairach space for neuroimaging was likely by Peter Fox in early PET studies. Results in this paper were reported via the use of co-ordinates to enable comparison between studies and sites. Only a very limited number of neuroimaging studies since this use the prescribed manual technique to map data to Talairach Space. A few automated techniques exist to identify the intended landmarks and perform a true Talairach transformation (e.g. ANFI).
Instead, a typical mapping to “Talairach Space” in the current literature is prescribed as an automatic registration to the average305 or ICBM152 models. To a novice user of these techniques one could be forgiven for thinking that the average 305 is indeed “Talairach Space”, given that it was constructed using the co-ordinate system of Talairach, unfortunately it isn’t and was never intended to be. In order to minimise ambiguity, a mapping such as this should be described as a mapping to “MNI space”.
Another point to note is that given that most automatic registration techniques use linear registration to map the whole cortex of a novel brain to MNI space, there will be an amount of residual variability remaining after the mapping especially in cortical surface areas in these studies. In contrast, a true Talairach mapping will also contain some variability in these areas but will be better defined around the AC-PC line.
This is not a reason to discard either of the techniques but rather a cautionary note regarding their usage, typically the benefits of using an automatic registration technique far outweigh the use of a manual technique, the most pertinent being that the automatic technique will have a consistent bias and should be 100% reproducible. In this case, digital average models of anatomy alleviate this problem, as data can be directly matched using automated techniques to models such as the average305 and ICBM152. Collins et al originally validated such automated registration techniques with respect to manual Talairach registration in 1992.
What follows is a chronological description of the various models that have been created at the MNI, their intended use, limitations and notable places where they have been used.
Average 250 T1 modelEdit
In 1992, Evans et al created the original average 250 model. It was constructed by manually identifying anatomical landmarks in 250 MRI scans of young individuals that were acquired as part of ongoing PET studies. These landmarks were chosen from the Talairach and Tournoux atlas and thus the final average and space approximates that of Talairach and Tournoux. The manually identified landmarks from each of the subjects’ data were fitted together via least squares linear regression that was designed to match the resulting AC-PC line to the original Talairach and Tournoux atlas. The resulting AC-PC best fit line matches that of Talairach but has a small residual bias in the z direction. This original MNI-space model has never been officially released from the MNI.
The native data from these acquisitions was 256x256 with 2mm slices. The final image resolution of this data was 172x220x156 with 1mm isotropic voxels.
Average 305 T1 modelEdit
A further 55 datasets that used the same scanning protocol were added to the original 250 resulting in 305 subjects (239 males, 66 females; age: 23.4 +/- 4.1 yr.). This data was then linearly fitted using a 9 parameter transformation to the original average 250 using an entirely automatic fitting strategy mritotal developed by Louis Collins as part his thesis dissertation (Error: Reference source not found). The resulting average is again only an approximation of the original Talairach space and in this case, the Z coordinate is approximately +3.5mm at the origin in MNI-space. The predominate reason for this difference is as mritotal is a whole-brain fitting strategy in deference to the Talairach approach.
The native data from these acquisitions was 256x256 with 2mm slices. Once resampled, this data resulted in the original average305 and has subsequently defined MNI-space. The final image resolution of this data was 172x220x156 with 1mm isotropic voxels.
Colin 27 T1 modelEdit
In 1998, a model with much higher definition was created in order to address the need for a model that approximated Talairach space with less local variability than is inherent in linear models such as the 250 and 305 models. The primary reason for this was for localisation of functional data to average space. To achieve this an individual (Colin Holmes) was scanned 27 times over a period of 3 months using multiple sequences (). The images were then all linearly registered to each other and an average created that exhibited very high SNR and structure definition. This average was then linearly registered to the average305 using mritotal. The Colin 27 model exhibits a closer match to the structures in the Talairach atlas, but is not perfect, for example; the origin is shifted approximately +3.0mm in Y and -4.6mm in Z as compared to the Talairach model.
The native and final resolutions of these images vary due to various sub-studies that were run as part of this scanning set but are largely and at least the same resolution as the average305.
ICBM 152 T1, T2 and PD modelsEdit
In 2001, within the context of the ICBM project a concerted effort was undertaken by three sites (MNI, UCLA, RIC) to collect a set of full-brain volumetric images from a normative population specifically for the purposes of generating a model. These images were acquired at a higher resolution than the original average 305 data and exhibit improved contrast due predominately to advances in imaging technology. Each individual was linearly registered to the average 305 and a new model was formed. In total, three models were created at the MNI, the ICBM152_T1, ICBM152_T2 and ICBM152_PD from 152 normal subjects. One advantage of this model is that it exhibits better contrast and better definition of the top of the brain and the bottom of the cerebellum due to the increased coverage during acquisition.
The entirely automatic analysis pipeline of this data also included grey/white matter segmentation via spatial priors (). The averaged results of these segmentations formed the first MNI parametric maps of grey and white matter. The maps were never made publicly available in isolation but have formed parts of other packages for some time including SPM, FSL AIR and as models of grey matter for EEG source location in VARETTA and BRAINWAVE.
Again, as these models are an approximation of Talairach space, there are differences in varying areas, to continue our use of origin shift as an example, the ICBM models are approximately +3.5mm in Z and +2.0mm in Y as compared to the original Talairach origin. In addition to the standard analysis performed on the ICBM data, 64 of the subjects data were segmented using model based segmentation. 64 of the original 305 were manually outlined and a resulting parametric VOI atlas built.
The native data from these acquisitions was 256x256 with 1mm slices. The final image resolution of this data was 181x217x181 with 1mm isotropic voxels.
Nonlinear ICBM 152 T1, T2 and PD modelsEdit
In 2002 a further refinement of the registration technique used to generate the initial linear ICBM models was performed by Andrew Janke and Louis Collins as part of building on existing recursive registration work of Alexandre Guimond. Here, each of the individuals in the ICBM population were iteratively matched to their own evolving internal average using non-linear registration. It was found that after six iterations of this technique that the model stabilised when using a 4mm fit. This technique allows a model to built that exhibits less of the anatomic variance that exists within the group and better express the commonality between the individuals. A model with better defined contrast also makes a better registration target than a simple linear model when using nonlinear registration to match an individual to a model such as the ICBM. Vladimir Fonov has since taken these models and further refined technique with better resampling and fit characteristics, this has resulted in the ICBM 2009 a, b and c averages that are currently released as the nonlinear MNI average.
In 1996 an effort was undertaken to better facilitate automated regional segmentation and structure identification via non-linear registration, a single subject was manually segmented into 142 sub-regions by trained neuroanatomists using a standardized nomenclature based on a specific hierarchical scheme. To increase the SNR, two averages were acquired from a subject who exhibited a relatively symmetric cortex. The objective of this atlas was to produce a completely space-filling atlas of a single MRI volume, where every voxel could be identified as part of brain or extra-brain structure. This atlas would serve as a template model for automatic segmentation of other MRI volumes and in the creation of a three dimensional probabilistic atlas of the human brain.
The creation of this atlas was inspired by a proposal put forth by the International Consortium of Brain Mapping (ICBM). The specific goal of the ICBM project was the development of a population-based digital probabilistic atlas of the human brain with a reference system in stereotaxic space. In order to produce such an atlas, a starting template was needed. Other than the advantage of being a teaching tool for normal anatomy, the probabilistic atlas may also be used for the localisation of activation sites in functional studies and as a standard against which pathological brains can be analysed.
In order to generate probability maps of the structures that had been traced on the individual, each of the 152 ICBM subjects were nonlinearly registered using ANIMAL to the traced dataset. This then allowed the labels from the representative subject to be transformed to each of the individuals. These labels on each of the 152 ICBM subjects were then linearly mapped to the ICBM space using the same linear transformations that were used to generate the initial average 305 target. In this case the average 305 target was used if only because the ICBM152 T1 model was still a work in progress. . Given that the resulting SPAMS (Statistical Probability Anatomic Maps) were expressed in ICBM space, they were provided to a few groups including the San Antonio Imaging Centre. This data was then used to form part of the original Talairach Daemon. The Talairach Daemon data was then subsequently inherently included in the WFU pickatlas that is typically included as an extension to SPM. The data that is included in the Talairach Daemon has since undergone a small amount of revision in order to better approximate the original Talairach space. The original parametric maps have not been made available.
This information was collected between 2006 and 2010 by Andrew Janke and to the best of my knowledge is a somewhat concise history of Atlases at the BIC, no doubt some information is missing and some perhaps wrong, if this is the case please email me (email@example.com) so that I can correct it or perhaps just edit it yourself.