B2-COS similarity

A similarity metric measures the “distance” between pairs of t-maps. Similarity, in Figure 1, is measured by B2COS (Ibraev, 2005). Thresholded t-maps (preserving 1% of all voxels) are projected to lateralized Brodmann regions (Brodmann, 1909) (each Brodmann area is divided into distinct left/right regions). This forms an 82-component vector in which each entry represents fraction of voxels in the corresponding Brodmann region that are “active” (i.e., above the threshold). The cosine of the angle between these vectors (B2COS) is used as the similarity metric in this figure. This snapshot contains a hit matrix summarizing the similarities among the fMRI datasets contained in the repository. Approximately 1000 experiments have been analyzed and are ordered along the X/Y axis (from right to left, top to bottom). Similarities, using the B2COS metric, are represented by color intensities (Yellow-high similarity, Blue – low similarity). Note how there are block based regions visible in the snapshot, such as the study experiments.

Figure 1:Similarity matrix using the B2-COS metric for all datasets in our repository. (click on the image to view the full size matrix)

Note: The image is quite large and Internet Explorer has trouble opening it.

If you cannot see the picture by clicking on the above link, you should save the image to your machine (right click -> Save Target As...) and open it with your default image viewer.

Concentrations of similarity are particularly strong for the “study” runs, and we see (although difficult to resolve at this scale) particularly strong similarity in the “study face” condition. Off diagonal concentrations of similarity represent a particular challenge, suggesting relations between apparently distinct cognitive experiments. For example, the clear off-diagonal lines in the “morality data” represent strong within subject similarities.

For quantitative assessment, we build the set of so-called “Receiver Operating Characteristic” (ROC) curves, treating each data set separately as a query, and scoring as a “true positive” the retrieval of a data set from the same experimental condition (examples at Figure 2(a), (b)).

Figure 2: (a) and (b) are the ROC curves for the study recall experiments. The ROC area histogram, which summarizes all the ROC curves, is shown in (c).

Each curve is summarized by computing the area under the ROC.If the retrieval is no better than random, this area is 50%.The histogram of the areas, Figure 2(c), provides a high level summary of the information represented in the visual matrix.Current results, using the B2COS measure show that most areas exceed 50%, and that whole set of areas seems to consist of a large distribution with mean that is somewhat better than random, and a long tail to the right of cases where this similarity measure works extremely well. The full list of ROC curves for every experiment in our repository can be seen here: