Title: RealTime Simulation and Visualization of Volumetric Brain Deformation from intraoperative MR images
1Real-Time Simulation and Visualization of
Volumetric Brain Deformation from intraoperative
MR images
- M. Ferrant, A. Nabavi, B.Macq,
- R. Kikinis and S.K.Warfield
2Context
- The context of our research is image-guided
neurosurgery (IGN) - During IGN, the brain has been observed to deform
mainly because of gravity, CSF draining/leakage,
anaesthesiology and tumor resection - It is therefore of crucial importance to be able
to correct for these deformations by updating
pre-operative imaging the surgical planning is
based upon accordingly. - Also, it is desirable for the neurosurgeon to
understand these deformations with appropriate
visualization tools.
3Previous work and goal
- In our previous work, we have developed a set of
image processing algorithms for tracking and
registering intraoperative MR images of the brain
during neurosurgery using a FE biomechanical
model. - We have optimized our algorithms so they can
execute the registration in a time compatible
with the constraints of neurosurgery. - In this paper, our goal was to efficiently
visualize and characterize the deformation of the
brain from intraoperative MRI sequences in real
time.
4Brain Shift During Surgery (1)
Before surgery (1)
Dura opened (2)
Small resection (3)
Tumor resected (4)
Dura closed (5)
5Brain shift during surgery (2)
Time T3 IOP image (resection has started)
Time T4 IOP image (resection progresses)
Time T2 IOP image (after removal of dura)
Time T1 Pre-operative image
Time T5 IOP image (dura closed)
? How to match image at time T1 onto image at
time Ti ?
6 Brain shift during surgery (2)
7Overview of deformable registration (1)
- We track landmark surfaces (lateral ventricles,
cortical surface, tumor if needed) in the image
sequence and infer the volumetric deformation
field from the surface deformations using a
biomechanical FE model. - Our algorithm consists of 4 main steps
- Generation of a FE tetrahedral model from the
initial image, with boundary surfaces as landmark
surfaces, - Deformation of the boundary surfaces onto
successive intraoperative images using an active
surface algorithm, - Sucessively infer a volumetric deformation field
using the active surface deformations as boundary
conditions for the initial FE mesh. - Update preoperative imaging with the volumetric
deformation field
8Overview of deformable registration (2)
9Visualization tools
- We implemented a set of visualization modules
within the VTK library (www.kitware.com) so as to
be able to efficiently visualize and characterize
deformation of the FE model throughout the image
sequence in 2D and 3D. - Cuts through image data are displayed using
texture mapping. - Surface models and volumetric models can be cut
along arbitrary planes and overlayed on
corresponding cuts through image data. - Surface models and volumetric models can be
color-coded with the intensity of the deformation
they undergo or other characterization
parameters. - Deformation can be visualized as 2D or 3D arrows,
or as tensor data. (e.g. ellipsoids).
10Pre- and intra-operative Segmentation
- The segmentation serves to extract the key object
the surgeon wishes to track during surgery. - Pre-operative segmentation is done accurately
with time-consuming algorithms - Intra-operative segmentation is done in real-time
using the pre-operative segmentation as a
patient-specific atlas.
11Intra-operative segmentations
12Construction of the FE model
- A multiresolution tetrahedral mesh is extracted
from the brain volume using an algorithm that we
specifically designed for generating FE meshes
from image data. Boundary surfaces can be
extracted from the volumetric mesh as
triangulated surfaces. - The algorithm is available for research purposes
13Visualization of surface deformation (1)
- The algorithm matches the preoperative landmark
surfaces iteratively onto the successive target
intraoperative images.
14Visualization of surface deformation (2)
After dural closing
15Visualization of surface deformation (3)
16Topological update of model (1)
- Elements covering the resected areas are removed
from the mesh and those lying across the brain
boundary are clipped
17Topological update of model (2)
- Cuts through topologically updated brain surface
on corresponding intraop scan.
18Computation of Biomechanical Simulation
- Estimate three displacements at each mesh vertex
of the volumetric FE mesh. - Active Surface algorithm provides known
displacements at some mesh vertices (those of the
boundary surfaces). - Sparse linear system of equations, depending on
the connectivity of each vertex in the mesh. - Resolution of the matrix system equal
distribution of equations to compute nodes of the
cluster. Assembly and resolution of matrix system
are parallelized using MPI and the PETSc library.
19Scaling on parallel cluster
- Timings of biomechanical simulation on Alpha
cluster with 16 nodes. - Mesh with about 125000 tetrahedra and 77511
Equations to solve.
20Visualization of volumetric deformation (1)
21Visualization of volumetric deformation (2)
22Visualization of volumetric deformation (3)
23Visualization of volumetric deformation (4)
3 TO 4
4 TO 5
24Updating preoperative imaging
- The volumetric deformation field can be
interpolated back onto the image grid to deform
the intraoperative image at the previous time
point.
25Landmark deformation analysis
- Landmarks were manually placed on slices of the
intraop sequence and deformed with the algorithm
for comparison. - Mean error was about 2.5 mm
26Characterization of deformation (1)
- We visualize stress-tensors using ellipsoids.
Color-coding is done with the largest eigenvalue
of the tensor.
27Characterization of deformation (2)
28Analysis Timeline
Typical times with current implementations on
current hardware.
29Summary
- We have developed and optimized 3D and 2D
visualization modules for the real-time
visualization of biomechanical deformation from
image sequences. - The visualization tools allow for improved
characterization, understanding and
visualization of brain deformation during
neurosurgery. - We have optimized our deformable registration
algorithm by parallelizing it so it runs within
the time constraints of neurosurgery.
30Future work
- Visualization tools
- plug all visualization modules into a single
program - ? Into the 3D slicer (D. Gering et al.,
MICCAI99) - Algorithmic
- improve surface matching algorithm
- include other structures in the model (e.g.
falx) - use more accurate constitutive material equations
(e.g. visco-elasticity) - run the algorithm on a large number of cases for
validation, as well as during surgery.
31Extreme brain shift !? ...