Indexing of brain images using various similarity metrics


We tried several different metrics for quantifying the similarity of two processed brain images (t-maps). All t-maps were thresholded so that only 1% of the brain voxels remained active.

Name and short description
Similarity matrix
ROC curves
B2-COS

= Brodmann left/right differentiated cosine.
Each t-map is converted into a vector or 82 components, each component representing the percentage of overlap between voxels in the t-map and one of the 82 left/right differentiated Brodmann regions.

Data: odd, ep, rec, mor, rom

Symmetric Difference EMD

The symmetric difference EMD computes the earth mover's distance between the symmetric difference of two Brodmann regions.

Data: rec, mor (partial)
...
Symmetric Hausdorff

The symmetric Hausdorff distance is the classical definition of the Hausdorff distance.

Data: rec, mor (partial)
...
Average Hausdorff

Modification to the Hausdorff which reduces the effect of outliers

Data: rec, mor (partial)
...
Simple Overlap

Similarity is defined as the percentage of overlapping voxels with respect to the total number of voxels in the two datasets.

Data: odd, ep, rec, mor, rom

B2-Overlap

Simple overlap averaged over the 82 left/right differentiated Brodmann areas.

Data: odd, ep, rec, mor, rom

C-COS

=cubes cosine.
The brain was partitioned into 74 axis alingned cubes of about the same size. Each t-map is converted into a vector or 74 components, each component representing the percentage of overlap between voxels in the t-map and one of the 74 cubic regions.

Data: odd, ep, rec, mor, rom
R-COS

= random cosine.
The brain was partitioned into 82 regions of about the same size by randomly assigning each voxel in the brian to one of the regions. Each t-map is converted into a vector or 82 components, each component representing the percentage of overlap between voxels in the t-map and one of the 82 random regions.

Data: odd, ep, rec, mor, rom
...



Last Updated: Aug10, 2005