Region of Interests ROI Extraction and Analysis in Indexing and Retrieval of Dynamic Brain Images

1
Region of Interests (ROI) Extraction and Analysis
in Indexing and Retrieval of Dynamic Brain
Images Researcher Xiaosong Yuan, Advisors Paul
B. Kantor and Deborah Silver Sponsored by
National Science Foundation (EIA-0205178)
1. Introduction In functional magnetic resonance
imaging (fMRI) time-series analysis, Region of
Interests (ROI) needs to be extracted to trace
the paths of activation that are relevant to
various activities of human brains. The fMRI
4-Dimensional datasets in this study comprise a
group of experiments with finger-tapping in
certain defined patterns. We present a fast and
effective method to extract the regions which are
related to the variance of blood flows in brains.
The method is based on the correlation function
between experimental stimuli and observation
signals with noise. The statistical significance
of the response to the stimulus pattern can be
obtained together with its time lags.
Fig. 1. Mean volume over time of the original
brain dataset (32 slices)
2. Finger tapping patterns Bimanual simultaneous
opposition of thumb and four fingers at a "fast
but comfortable rate" (approx. 3-6 Hz)).
paradigm 16 blocks, each 24s, tapping, rest,
tapping, rest, . scanning parameters 132
acquisitions, TR3s, TE30ms, 32 axial slices.
Fig. 2. The intensity of the observation signal
over time. The six voxels shown are with high
correlation values
Fig. 3. The distribution of the time lag in
baseline (right) and finger-tapping (left)
Fig. 4. The distribution of the time lag of the
voxels in the brain with high confidence levels
Fig. 5. The Cross-correlation Rxy(l) at maximum
Fig. 6. The test statistic t value histogram
(left) and volume map (right) from the
correlation Rxy(l)
4. How to generate the brain mask volume The mask
(1/0 binary map) is used to classify all voxels
into two groups - inside or outside of the brain.
So that in the future processes we can tell
whether a voxel is a brain voxel or a background
voxel, such as drawing the histogram of the brain
volumes, etc. The intensities inside of the brain
are almost always higher than the background,
therefore we can threshold for it. This will
generate a binary image (mask0). The mask0 has
both holes inside and islands outside the brain.
Therefore we want to take away these holes and
islands to make a solid sphere mask. The mask
process consists of two steps Step 1. Get rid of
the holes from mask0 first (1) Pick up a seed at
(0, 0, 0), flood-fill the outside space with
certain value k. (2) transform every voxel with k
into 0 every voxel not with k into 1. (3) Thus
we get the mask1 only with a solid sphere and the
islands.
Fig. 7. The Mask0 and final Mask generated
Step 2. Get rid of the islands from mask1 (1)
Pick up a seed at the center point of the brain ,
flood-fill the inside space with certain value
k. (2) transform every voxel with k into 1 every
voxel not with k into 0. (3) Thus we get the
final mask only with a solid sphere.
5. Conclusion We present a method for identifying
brain regions that show a lagged correlation with
the presumed stimulus. This seems to be a
promising alternative to the standard Statistical
Parametric Mapping method. In particular, this
approach gives each identified brain region a
time lag parameter, which may make it possible
to track an activation through the brain over
time. The results above show that the method does
not produce artifacts when there is no
interesting task, the histogram of correlations
shows no interesting features.
March 11, 2003 in APLab, SCILS