OBJECTIVES

1

• Prior work has demonstrated that anatomical
  variability is too great for standard AC-PC
  registration techniques to produce meaningful
  clustering.
• This is the first 3-D database of population
  neurophysiologic observations that clearly
  demonstrates

INTRODUCTION
METHOD
RESULTS 89 Patients, 107 Procedures, gt 15
000 data points
Visualization Software
Data Clustering in Reference Brain Space
We present a method of constructing a database of
intra-operatively observed human subcortical
electrophysiology. In this approach, patient
electrophysiological data are standardized using
a multi-parameter coding system, annotated to
their respective magnetic resonance images
(MRIs), and nonlinearly registered to a
high-resolution MRI reference brain. Once
registered, we are able to demonstrate clustering
of like inter-patient physiologic responses
within the thalamus, globus pallidus, subthalamic
nucleus, and adjacent structures. These data
may in turn be registered to a 3D patient MRI
within our image guided visualization program
enabling prior to surgery the delineation of
surgical targets, anatomy with high probability
of containing specific cell types, and functional
borders. The functional data were obtained from
89 patients (107 procedures) via microelectrode
recording and electrical stimulation performed
during stereotactic functional neurosurgery at
the London Health Sciences Centre (London,
Ontario, Canada).
Visualizing complex three-dimensional
somatotopography required the construction of a
novel image-guided neurosurgery program. The ASP
(Atamai Surgical Planning 1) platform allows
interactive display of 3-D objects, planning of
surgical trajectories and visualization of
stimulation and microelectrode recording (MER)
derived functional data.
1. Axial image at base of thalami. Green
Spheres 585 kinesthetic data points acquired
using MER in 28 patients from the Vim nucleus.
Blue spheres 375 tactile data codes obtained
using MER in 22 patients. Note naturally
evolving probabilistic functional border between
the Vim and Vc nuclei.
2. Obliquely oriented axial image positioned at
level of mid thalamus. Bilateral representation
of tremor synchronous cells detected using MER
are displayed as 376 red spheres. Data were
obtained from 35 patients. The left thalamus has
been segmented and the top aspect of the anatomy
removed to reveal contents.
Acquire and Code Patient Data
Patient data (MRI and intra-operative records)
from patients who received stereotactic
neurosurgical treatment for chronic pain and
movement disorders were obtained from the London
Health Sciences Centre (LHSC), London, Ontario,
Canada.
3. Clustering of left subthalamic nucleus data
codes. Larger image provides context and
displays trajectory of a quadripolar dbs
electrode inserted into left STN. Inset shows
functional data acquired from 3 patients during
STN dbs surgery clustered around contacts of
electrode. Shown here are tremor synchronous
cells, kinesthetic cells, cells with firing
patterns typical of STN neurons, and larger
spheres encoding for stimulation-induced
paresthesias.
All intra-operative data were quantified and
standardized using a GUI integrated into ASP. A
six-parameter code was assigned to every data
point collected from each patient containing the
following information
OBJECTIVES
Resulting Code
Nonlinear Versus AC-PC Based Registration
069_1_19?A_R_12_PT
Images display a segmented portion of the visual
pathway. Included are proximal portions of the
left and right optic nerve, optic tracts and the
optic chiasm (OC). The left optic nerve and
tract are labeled (LON, LOT).

• This research addresses the following questions
• Will a database of population electrophysiologic
  observations present meaningful clusters of like
  data?
• Will a fast, unsupervised nonlinear registration
  algorithm produce tighter clusters of like data
  over standard AC-PC registration techniques?
• Will the database predict the placement of
  lesions prior to surgery?

Annotate Coded Data to Respective Patient MRI
During the coding procedure, standardized patient
data were retrospectively annotated to their
respective MRIs at the patient image coordinates
where they were evoked. The software
automatically calculates the Leksell-to-Image
coordinate transformation.
Both images display the same 153 data codes which
represent microstimulation induced visual
responses described by 28 patients during
pallidotomy (typically as flashing lights).
The data on the left (red spheres) were
registered to the reference image using standard
AC-PC fitting techniques. On the right, data
(white spheres) were registered to the reference
brain using our nonlinear registration algorithm.
Both images include insets which display the LOT
as a wireframe mesh so stimulation spheres
internal to the segmentation may be seen. Note
the visibly tighter clustering provided by
nonlinear registration. Sphere diameter
represents current magnitude at 300Hz, 0.2 ms
pulse.
A typical left thalamotomy trajectory Purple
Electrode Cyan Trajectory Cross Hairs Zero
mark on trajectory Yellow Lines AC-PC planning
software Spheres indicate stimulation-induced
responses and are scaled to represent amount of
current employed to evoke the response. All
points collected at 300Hz, 0.2ms pulse, with a
0.12mm diameter tungsten electrode.
ADVANTAGES OF THIS APPROACH
A nonlinear registration algorithm accommodates
for inter-patient anatomical variability during
database-to-patient registration. We have
developed a nonlinear registration algorithm
which matches, as closely as possible, the
anatomy of a patient MR image to a high
resolution, high contrast to noise, reference
brain MRI. The inverse of this registration is
used to map the contents of the database from
reference brain image space back into patient
image space. Digitized versions of atlases
of anatomy are not used as the common database
coordinate system. Historically, patient
functional data were scaled to versions of
anatomic maps so population functional
organization could be studied in relation to
anatomic structures. However, plotting
inherently three-dimensional data on a
two-dimensional atlas necessitates a 3-D to 2-D
compression of patient data to fit the atlas
slice considered most representative of the
volume of brain being explored. Using the high
resolution reference brain MRI as the coordinate
system for the database allows us to avoid the
use of digitized anatomical atlases
altogether. Advanced data mining
capabilities. Graphical user interfaces coupled
to a search engine make generic or detailed
searches of the database possible. The user may
extract and display only those data most closely
approximating the current patients age, sex,
handedness, and diagnosis, or conversely, choose
to view only data representative of a larger
cross section of the database population. The
database is a constantly evolving tool. With the
analysis of each new patient, the database grows
in size with results that may be immediately
appreciated. A digital probabilistic atlas of
this nature that utilizes population data will
improve in accuracy over time and achieve better
statistics upon the addition of each new
patients data.
Post-operative Outcome Lesion Analysis
Patient 015 Left thalamotomy for essential
tremor which produced complete resolution of
contralateral tremor at 7 month follow-up. Left
Axial image placed at base of lesion. Right
Magnification of lesion (wireframe
representation).
Nonlinear Registration of Data to Reference Brain
Image Space
Once tagged to the patient MRI, both the image
and the data points are nonlinearly registered to
a reference brain image. The reference brain is
a high-resolution, T1-weighted MRI consisting of
27 scans of the same individual averaged into one
volume 2.
Patient 054 Right thalamotomy for essential
tremor. Contralateral tremor returned after 2
months. Left Axial image placed at base of
lesion. Right Magnification of lesion
(wireframe).
We have developed a rapid and unsupervised
nonlinear registration algorithm, implemented in
vtk (www.kitware.com) and python
(www.python.org), that can compute each warp in
less than 10 minutes on a 1 Ghz Pentium III
laptop (256x256x248 MR image volume). The
methodology is comprised of two steps. The first
generates a global affine transformation that
maximizes the normalized cross-correlation
between the source and target (reference brain)
image. The second step builds upon the affine
registration by computing a deformation grid that
maximizes the similarity metric on successively
smaller sub-volumes of the images. The second
step may be constrained within any masked
sub-region of either volume.
In both patients, tremor synchronous cells are
shown, as is a 2D probability map indicating
where microstimulation induced tremor arrest in
20 patients. Red regions indicate greatest
probability for inducing tremor arrest, blue the
lowest. Note the poor outcome lesion overlaps
few of the desired database codes and does not
infringe upon the tremor arrest high probability
zone. The good outcome lesion is situated in the
densest region of each response type.
Lesion Content Breakdown Graphs present the
population data overlapped by each lesion shown
above once the database was nonlinearly
registered to the patient MRI. Bars represent the
percentage of total possible response codes
overlapped by the lesion. Green bars
microstimulation evoked, blue bars MER.
REFERENCES
1 www.atamai.com 2 .Holmes C.J., Hoge R.,
Collins D.L., Woods R., Toga A.W., and Evans A.C.
Enhancement of MR images using registration for
signal averaging. JCAT 22324-333, 1998.
Blue ellipse indicates warp mask. MRI voxels
outside of this 3D ellipse are ignored by the
algorithm. Cross-Correlation Map Red indicates
very poor correlation between target and source,
while transparent regions of the map indicate
excellent image correlation.