Transcript Slide 1

Nicu D.
1
Cornea ,
Dr. Ulukbek
1
Ibraev ,
A Visualization Tool for fMRI Data Mining
Prof. Deborah
1
Silver ,
1Rutgers
Database of fMRI data:
raw + analyzed
Query-by-example
Find similar
datasets
…
Investigating similarity
reported by other
methods
Data Mining
Results
…
Analyzed datasets
(activation clusters)
Prof. Paul
University,
2Drexel
1
Kantor ,
Prof. Ali
University,
2
Shokoufandeh ,
3University
How, Why
are these
similar ?
Cluster Comparison and Visualization Tool (CCVT)
Front end for data mining / content-based data retrieval engine
Prof. Sven
3
Dickinson
of Toronto
fMRI
Functional Magnetic Resonance Imaging (fMRI) is an increasingly popular imaging technique used to understand brain functionality. Scans of the subject’s head are taken at regular
intervals as the subject performs some mental task resulting in hundreds of 3D datasets. Large databases containing thousands of fMRI scans are already accessible to the research
community: the Brain Image Database (BRAID), the fMRI Data Center (fMRIDC), etc.
Motivation
• Ever increasing number of fMRI analysis tools and methodologies
• Difficult to compare their results
• many analysis parameters and various output formats
• In a database environment
• Want the ability to search for functional similarities in brain activation
• would permit new understanding of brain psychology
• Once similarity with another dataset is established
• Want to further investigate the reason for the similarity
• using other similarity metrics
Features
...
Jeff
2
Abrahamson ,
• Interactively query the similarity of processed (analyzed) fMRI datasets
• fMRI analysis result viewed as set of activation clusters
• cluster = a set of voxels grouped together by the analysis tool
• functional cluster
• statistical maps can be interactively thresholded
• comparison between multiple fMRI analysis results
• Two linked views of similarity
• Quantitative – table with inter-cluster similarity scores (similarity table)
• each cluster is mapped to a row and/or column in the table
• global view of similarity table in the form of color-coded bitmap
• Qualitative – interactive visualization of clusters in common brain space
• common area of selected clusters is highlighted
• Multiple similarity metrics
• overlap, nearest neighbor, etc.
• Query tools - formulate queries on the similarity table. Examples:
• filter table using user-supplied threshold
• show all pairs of similar clusters in two or more datasets
• for a specific cluster in one of the datasets, show all similar clusters in all
other datasets
• given a set of clusters in one dataset, show all clusters in other datasets
that are similar to ANY/ALL/UNION of the clusters in the selected set
Identifying regions activated in several
subjects of the same experiment.
A common activation area among 4 subjects of
an event perception experiment is identified by
filtering the table to remove similarity scores
smaller than 4%. Cluster A8 (of dataset A) shows
similarity with one other cluster from each of the
other datasets (B, C and D). The visualization
panel shows the overlap (in green) of A8 with
cluster D28 of dataset D.
Mapping analysis results to a brain atlas (Brodmann Regions)
The subject shows activation overlapping with Brodmann regions 9, 10, 11, 20,
21, 23 and 39. The similarity scores are above 5%.
Query-by-example data retrieval
Query dataset (A): study face condition (SFace), subject 7, is
more similar to the other SFace conditions (sets D, F and I).
Also, note that the similarity scores in columns labeled F (F1
and F4) are lower than those in the other columns. Set F
corresponds to the same condition, SFace, but performed by a
different subject (subject 4). Thus, in this example, we can
distinguish between datasets corresponding to different
conditions, and among those, we can differentiate between
different subjects, all based on the scores presented in the
similarity table.
Investigating similarity reported by other methods
(Brodmann vector: http://www.scils.rutgers.edu/~brim/PUBLIC)
each dataset is converted into an 82-component vector
representing the overlap with each of the 82 lateralized
Brodmann areas.
In this example, two datasets that show high Brodmann vector
similarity are compared.
Only 11 pairs of clusters out of the total of 1150 (25 x 46)
show any overlap at all; (less than 1% of the total number of
pairs). This demonstrates that the high Brodmann vector
similarity score is only partially due to actual voxel overlap
between the two datasets. The rest could be accounted for by
the different clusters that do not overlap with each other but
are in the same Brodmann areas.
Contact us:
E-mail: Nicu D. Cornea: [email protected], Prof. Deborah Silver: [email protected], Prof. Paul Kantor: [email protected]
Cluster Comparison and Visualization Tool home page: http://www.caip.rutgers.edu/~cornea/CCVT/