Title: MultiResolution Parameterization of Meshes for Improved Surface Based Registration
1Multi-Resolution Parameterization of Meshes for
Improved Surface Based Registration
b
a
a
a
- Sylvain Jaume - Matthieu Ferrant - Simon Warfield
- Benoît Macq - TELE - Université catholique de Louvain (Belgium)
- SPL - Harvard Medical School (USA)
a
b
2Surface Based Registration
- Surface Registration Matching a surface onto
object in 3D image
Applications
- brain atlas matching for automatic
labeling of sulci
- tracking heart boundary in real-time
MRI
- segmenting anatomical structures
3Common Approach Active Surface
- Active surface surface model deformed to the
object
boundary under some forces
(Cohen Cohen 93)
- Image force to attract the surface to the object
- Balloon force to increase capture range
- Internal force to limit bending
Limitation
Trade-off local accuracy
global smoothness
4Our Approach Multi-Resolution (M-R) Surface
Matching
- Build a Multi-Resolution reference surface of
the object
- Re-use the parameters of this surface to match a
new object
5Building M-R Reference Surface (I)
- External force move nodes x to minimize distance
x
2
T
2
Dist(x,T )
Area(T )
...
T
?
i
i
1
Dist(x)
x
x
?
1
Area(T )
T
i
0
j
j
x
distance to object 3D euclidean distance map
0
- Internal force smooth nodes to keep regularity
?
(x - x )
x
x
?
i
j
i1
i
j
6Building M-R Reference Surface (II)
- External force along surface normal or opposite
direction
Image gradient
Surface normal
- Internal force along this normal axis
.
x - x
i
j
(x - x )
x
x
n
n
i
j
i1
i
Parameters length of the deformation vector
7Building M-R Reference Surface (III)
- Refine by triangle quadrisection from resolution
i to i1
8Cuts through the Reference Surface
9Matching Surface onto a New Target Image
i1
Refine by
Surface at
Deform at res. i
Matched Surface
onto target
at resolution in
quadrisection
resolution i0
Reference
Extract parameters
Model
at resolution i
Potential function distance map of segmented
target image
10Cuts through the Matched Surface
Morphological closing was needed to avoid
flattening
11Conclusions of our M-R Surface Matching
- internal external forces are more more
localized
- convergence and accuracy of matched surface
- matching imperfectly segmented images
- difficult balance between int. ext. forces
- better potential function needed
- oversampling in smooth areas
Other benefits
- progressive transmission of the mesh