Fig.1: Examples of the comprehensive white matter atlas based on the
“Eve” single-subject atlas. In (D), the entire brain is parcellated
to 56 core and 46 peripheral white matter regions, in addition to 10 subecortical
gray matters, hippocampi, and others (total 130 structures, bilateral).
For the definition of peripheral white matter, cortex and the white matter
are not separated. In (E), the peripheral white matter is further segmented
from the associated cortex.
In this site, we offer several atlases with
different data sources and different image processing. Our atlases are still
evolving and they may cause some confusion. We hope the following explanation
would help you to understand how the atlases were created and should be
1) Single-subject or group-averaged:
i. Our original atlas was used in Wakana et al, Radiology, 230, 77-87
(2004) and “MRI Atlas of Human White Matter”, Elsevier. The
subject is a 22 years old male. This was originally called as JHU-DTI
single-subject atlas (a.k.a. “Adam atlas”). The matrix was
256x256x55 with FOV of 246x246x121 mm. It was not B0-corrected and did
not include the entire cerebellum. There were two re-sampled versions:
1 mm native space and Talairach space (both have 246x246x121 matrix).
ii. Our new single-subject atlas is now called JHU-DTI-SS (a.k.a. “Eve
atlas). The subject is a 32 years old female. This atlas covers the entire
cerebellum. The matrix is 181x217x181 with 1 mm resolution. This atlas
is B0-distortion corrected by non-linearly warping DTI data to a co-registered
T2-weighted anatomical image (Huang, Hua et al. 2005). There are MNI and
Talairach versions. The MNI version is linearly normalized to ICBM-152
template (MNI space) and the Talairach version was non-linearly normalized
to Talariach space by placing more than 300 anatomical landmarks read
from the original Talairach atlas (Talairach and Tournoux 1988).
i. Our current group-averaged maps are based on linear registration using
affine transformation of Automated Image Registration (AIR) (Woods, Cherry
et al. 1992).
ii. There are LBAM/JHU (JHU-DTI-GA) and ICBM (ICBM-DTI-81) versions. The
difference is subtle, mostly due to the difference caused by the phase
encoding direction and consequent difference in B0-distortion.
iii. Non-linear group-averaged maps are under construction. This atlas
uses the Eve atlas as the target.
2) The source of the data for the group-averaged atlases: Our atlas is
from either LBAM/JHU (https://www.mristudio.org/wiki/installation) or
ICBM database (http://www.loni.ucla.edu/Atlases/Atlas_Detail.jsp?atlas_id=15).
They were both acquired by 1.5T scanners but with different manufacturers
(JHU: Philips, ICBM: Siemens). Imaging parameters (96 x 96, 2.5 mm isotropic,
30 orientations) are similar but important differences are;
a. b = 700 s/mm2 for JHU and 1,000 s/mm2 for ICBM (consequently longer
b. The directions of phase-encoding and B0 distortion are different; JHU
= “exploded” frontal lobe vs ICBM = “imploded”.
3) B0 correction: B0 distortion correction was not applied to most of
our atlases. For linear-normalized maps such as JHU-DTI-GA or ICBM-DTI-81,
majority of the B0-distortion still remains. This means, the brains still
look like “exploded (JHU)” or “imploded (ICBM)”.
Effect of B0 depends on studies. If one is using only linear (or low-order
non-linear) transformation and is interested in intensity measurements
such as FA and ADC, B0-distortion can be treated as a part of cross-subject
anatomical differences. Here our interest is to align anatomy as much
as possible. If your data are not B0-corrected and you will use linear
transformation, you may want to choose JHU-DTI-GA for “exploded”
data and ICBM-DTI-81 for “imploded” data, although differences
are subtle (mostly in the frontal lobe) between the two.
4) Processed atlases: There are several “processed atlases”
based on the atlases described above. These processed atlases can be superimposed
on normalized subject data and used as automated 3D ROI analyses.
a. White Matter Parcellation Map (WMPM): These are hand-segmented maps,
in which known white matter structures are manually segmented based on
FA and color (fiber orientation) information. Peripheral white matter
regions (beneath the cortex) that are difficult to define manually were
defined by white matter probability in some atlases as described below.
In this case, we first define the white matter using FA threshold (0.25)
in each subject and normalized to a template. From these maps, white matter
probability at each pixel can be defined in the group-averaged maps. By
using threshold such as 0.9, reproducible white matter structures beneath
the cortex were defined (see Fig. 1a).
There are WMPMs defined in the group-averaged and single-subject atlases.
1. In the group-averaged atlases, structures that are highly variable
among normal subjects (mostly peripheral white matter regions close to
the cortex) are smeared out. The “core white matter” labeling
is limited to the well-known white matter regions around subcortical gray
matter. 50 structures are defined in these maps (Fig. 1b). (Mori, Oishi
et al. 2008)
2. For the peripheral white matter regions, we defined reproducible structures
using the probability threshold and identified 9 blade-type structures
beneath the cortical gyri. These structures called blades are also defined
in the WMPM (Fig. 1c).
3. In the single-subject atlases, all structures are sharply defined.
Our single-subject WMPM (Eve atlas) contains the same 50 core white matter
regions defined in the group-averaged atlases. The peripheral white matter
regions close to the cortex are also defined based on the group-averaged
results. There are several versions for the definition of the peripheral
a. A version that defines the entire cortex and underlying white matter
(46 regions) (Fig. 1d).
b. A version that separates the cortex and the underlying white matter.
This is the most comprehensive segmented atlas with 176 regions but it
contains many subject-specific anatomy and may not be suitable to be used
as automated segmentation (Fig. 1e).
c. A version that defines the peripheral white matter based on probability
(similar to the way the PWM was defined in the group-averaged maps but
this time using non-linear transformation) (Fig. 1f). In this map, cortical
parcellation is not provided.
b.Tract-probability maps: This map is created by averaging tractography
results in each subject. The tractography results performed in the native
subject spaces were first binarized into 1/0 information and normalized
to a template. By simply superimposing these binary data from multiple
subjects, tract probability can be calculated for each pixel (Hua, Zhang
et al. 2008).
c. Conceptual difference between the WMPM and tract-probability maps:
These two maps describe white matter anatomy with different concepts.
This is somewhat similar to state boundaries and roads in geological maps.
The WMPMs correspond to boundaries of states while the tract-probability
maps correspond to inter-state road maps. The WMPM can be used as an alternative
of classical ROI-based white matter delineation such as the manually defined
internal capsule. If one is interested in the status of a specific white
matter tract, such as the corticospinal tract (CST), the tract-probability
maps can be used. These two approaches may lead to different results.
For example, the tractography-defined CST occupies a portion of the internal
capsule. If a lesion, such as loss of FA, is localized only in the CST
portion of the internal capsule, the latter approach would give higher
sensitivity. However, there are several important points that should be
understood to use this approach. First, only a portion of the white matter
can be studied by this method. Second, if one finds that the CST has low
FA, it doesn’t immediately tell the CST is selectively affected.
It is possible that the entire internal capsule has abnormally low FA.
Therefore, these two approaches should be considered as complementary
5) Linear or non-linear transformation for automated white matter segmentation:
a. By normalizing subject brains to the atlas, or transforming the atlas
to subject brains, the white matter can be automatically segmented. This
is a powerful approach but it’s limitation needs to be well understood.
b. Currently all group-averaged maps were created by linear transformation,
which can adjust only the overall brain shape (height, length, and width).
Our group-averaged atlases are ideal as a target for linear or low-order
non-linear transformation because the atlases locate at the center of
the population. They are close to everybody. However, of course, there
could remain a substantial amount of mismatch between the normalized subject
images and the atlas especially for pathological brains. For example,
if the corpus callosum segment of the WMPM is applied to a patient group,
it may consistently miss the patient corpus callosum due to enlarged ventricles,
which would lead to low FA values. In this case, the analysis is detecting
consistent anatomical difference but not actual FA difference. The current
linear-transformation-based group-averaged maps and associated processed
atlas should be used as initial screening and interpretation of the results
c. For more precise anatomical analyses, non-linear transformation is
i. Non-linear transformation is sensitive to initial value. It is advised
to use linear transformation first to bring the subject data as close
as possible to the template before non-linear transformation. In this
sense, we advise to avoid Talairach atlas as a template because the brain
shape is substantially far from normal population.
ii. As a target of non-linear transformation, our experience suggests
single-subject atlas should be used. Very high-order non-linear transformation
could be confused by blur anatomical definition in group-averaged atlases.
iii. For the PWM segmentation, we recommend to use WMPM shown in Fig.
1c (linear or low-order non-linear transformation) or 1f (high-order non-linear
registration). An alternative approach is to apply the atlas shown in
Fig. 1d to each subject and segment the white matter within each cortical
segmentation manually or by thresholding (e.g. FA > 0.2).
||Fig. 1b and 1c
||Fig. 1d – 1f
||X (JHU data)
||X (ICBM data)
||X (ICBM data)
Atlas-based whole brain white matter analysis using large deformation diffeomorphic metric mapping: application to normal elderly and Alzheimer's disease participants.Oishi K, Faria A, Jiang H, Li X, Akhter K, Zhang J, Hsu JT, Miller MI, van Zijl PC, Albert M, Lyketsos CG, Woods R, Toga AW, Pike GB, Rosa-Neto P, Evans A, Mazziotta J, Mori S. Neuroimage. 2009 Jun;46(2):486-99.
Human brain white matter atlas: identification and assignment of common anatomical structures in superficial white matter.Oishi K, Zilles K, Amunts K, Faria A, Jiang H, Li X, Akhter K, Hua K, Woods R, Toga AW, Pike GB, Rosa-Neto P, Evans A, Zhang J, Huang H, Miller MI, van Zijl PC, Mazziotta J, Mori S. Neuroimage. 2008 Nov 15;43(3):447-57. Epub 2008 Jul 18.
Hua, K., J. Zhang, et al. (2008). "Tract probability maps in stereotaxic
spaces: Analyses of white matter anatomy and tract-specific quantification."
Neuroimage 39(1): 336-47.
Huang, H., K. Hua, et al. (2005). Characterization and correction of
B0-susceptibility distortion in SENSE single-shot EPI-based DWI using
manual landmark placement. ISMRM, Miami.
Mori, S., K. Oishi, et al. (2008). "Stereotaxic white matter atlas
based on diffusion tensor imaging in an ICBM template." Neuroimage
Talairach, J. and P. Tournoux (1988). Co-planar stereotaxic atlas of
the human brain.
3-Dimensional proportional system: An approach to cerebral imaging. New
York, NY, Thieme Medical.
Woods, R. P., S. R. Cherry, et al. (1992). "Rapid automated algorithm
for aligning and reslicing PET images." J.Comput.Assist.Tomogr. 16: