The Online Brain Atlas Reconciliation Tool (OBART) provides an interactive method to examine quantitative relationships between brain regions defined by different digital atlases or parcellation methods. OBART's current focus is for human brain imaging, though our techniques generalize to other domains. The methodology and some results were presented in Bohland et al. (2009). The brain atlas concordance problem: quantitative comparison of anatomical parcellations. PLoS ONE. 4(9): e7200.


The OBART database contains spatial relationships between pairs of brain regions defined by different atlases. In particular, OBART compares different labelings of the single subject MNI template brain at 1mm resolution. Labelings are compared for the Left Hemisphere (LH) only due to the fact that some atlases explicitly distinguish LH from RH structures, while others do not. For all labelings, only LH voxels of the template brain which are likely to be gray matter are included; these voxels are selected by thresholding the output of SPM5 software's probabilistic tissue classification (gray matter output) at a value of 0.25.

For a region i selected from one atlas, and a region j selected from another atlas, the value Pij represents the region pair's asymmetric conditional overlap:

where ri represents the set of all voxels assigned that label, and where x indexes voxels. This measure has at least two simple interpretations:

  1. It is the fraction of region j that is contained within region i. Likewise, Pji then becomes the fraction of region i contained within region j.
  2. It can be considered to be a conditional probability. If we know that a voxel is in region j (and know nothing more than that, as may be the case when reading the published literature, for example), then Pij indicates the probability that the voxel also falls within region i.

In addition, a symmetric (Composite) measure can be computed from the pair Pij and Pji as follows, indicating the amount of overlap normalized by the geometric mean of the two regions:

Explore Atlases

The above described quantities are available for all region pairs across a number of diverse labeling methods applied to the single subject template brain. Starting HERE you can select one atlas (see left, below), then (from the hyperlinked page) select a region from that atlas, then determine what other regions (from a set of other atlases / labelings of your choice) overlap the chosen region (and to what extent using the measures above).

An example of a portion of the results page for is seen below for the region Broca's Area BA44 from the Anatomy Toolbox.

Various metadata including projection images showing the region definition are shown. Each term has been mapped to a standardized table and - where applicable - a linkout is provided to the appropriate category within the NeuroLex ontology.

Compare Atlases

OBART also provides visual representations that allow the user to examine region to region correspondences for a chosen pair of atlases / labelings. Such comparisons can be peformed by choosing two atlases from the entry page HERE, and choosing the Compare two atlases button (see below). These quantitative comparisons are again based on labeling of the MNI single subject template (Left Hemisphere only).

For each atlas pair, we construct a bipartite graph in which vertices (nodes) represent individual brain regions, with one set representing the first chosen atlas, and the second set the second chosen atlas. The vertices are represented as two vertically stacked sets, with edges drawn only from one set to the other. These edges correspond to spatial overlap relations. In particular, each edge has a weight:

where V1 and V2 are the vertex sets (regions) comprising each atlas. Thus the edge weight will be zero for pairs of regions that do not overlap, and 1, for example, when a region is a pure sub-region of another. Intermediate values represent varying degrees of partial overlaps.

OBART displays the bipartite graph images for a user-specified edge threshold. By setting this threshold (using the slider control or edit box at top, and then clicking the Refresh Threshold button), OBART will remove all edges with weights Eij with values less than the chosen value. Conceptually this means that partial overlap values below your chosen limit will be removed from the representation. This can be thought of as a noise reduction process in that a value of, for example, Eij indicates that only 10% of one region is contained within the other, and you may consider this to not indicate a strong region-to-region correspondence. The default threshodl value is 25% (or 0.25), but exploration of different thresholds leads to insight into overall correspondences between the regions in two atlases. The two images below show correspondences between two atlases at different thresholds (10% and 25%).

OBART calculates the connected components in the graph after thresholding. These are color coded to indicate sets of regions in the two atlases with approximate spatial correspondence (as a whole). For example, in the image above (25% threshold), the first connected component (at top) shows that the precentral gyrus in the LPBA atlas is approximated by the union of the precentral lobule, rolandic operculum, and precentral region in the AAL atlas. You can go directly to these graph visualizations with the chosen thresholds by clicking on the images above. As the threshold is increased, the graph will break into more and more connected components, but be careful in your interpretation as you are removing real correspondences to arrive at a more simplistic representation.

Multi-atlas Labeling Tool

The labeling tool provides the ability to obatin brain region labels and corresponding spatial overlaps between all regions represented in the OBART database and a region mask that you provide. The system allows the user to upload her own NifTI-1 formatted data volume (or gzipped NifTI volume), which has already been registered to MNI-305 space.

Your volume should be a "mask" of a brain area or areas, with volume entries set to 1 (or any non-zero value) for voxels to be included in the mask, and 0 for voxels to be left out of the mask (background). OBART will then perform the same overlap calculations as described above, treating your mask as a new single "brain region" to be compared against all others, and will return results in a tabular summary as above.


The OBART project is funded by generous support from the NIMH (5R01MH084802, PI Partha Mitra).