Tweet
Please find below a discussion on mesh generation techniques from 3D
imaging data which I hope will be useful.
I must state that I am one of the founding members of Simpleware Ltd
which markets the software package ScanFE for the generation of finite
element (and CFD) meshes from 3D image data - however I am also an
academic with a long standing interest in this area and I would like to
think that the comments below provide, as far as possible, an impartial
assessment of the different routes available to converting CT data to FE
mesh.
There are essentially two distinct approaches to the generation of
finite element meshes from 3D imaging data (as obtained from MRI, CT,
micro-CT for example): CAD based and "voxel" based approaches.
Both approaches require the segmentation of the 3D data; in other words
the identification of volumes of interest within the image.
Step 1: Segmentation
Segmentation techniques can range from simply defining a threshold (e.g.
anything above a certain greyscale is bone) to sophisticated
combinations of complex image processing algorithms. Segmented images
consist of one or more volumes of interest (flagged voxels within the 3D
image representing bone, ligaments, muscle, fat..) which are often
called masks. Segmentation tools are provided by a wide range of
software packages including, in addition to those listed in previous
emails, Analyze (from Mayo clinic) and ScanFE.
Once masks/volumes of interest have been obtained using segmentation
tools two distinct routes to generating meshes are possible.
"CAD based" approaches
Step 2 for "CAD based" approaches: convert the segmented masks (volumes
of interest) into geometric surface representations bounding the volumes
of interest: examples of such representations include tessellated
surfaces (described by primitives such as triangles), stacks of contours
which then need to be lofted into surfaces, or NURBS type patches
(higher order polynomial representations) of the boundaries. The
conversion from volume data to surface representation is often fraught
with difficulties:
PROBLEMS
i) gaps frequently appear in surface description which
need to be painstakingly corrected manually,
ii) where two or more volumes are in contact their surface
description are usually non-conforming (gaps/overlaps)- this is a
significant problem for modeling interfaces particularly contact
surfaces,
iii) stacking two dimensional contours (essentially a so-called
21/2 D approach) is a particularly poor approach as it cannot
satisfactorily/automatically handle bifurcations
iv) the generation of NURBS surfaces (and equally the contour
stacking approach) almost inevitably engenders approximations and the
loss of features.
Step 3 for "CAD based" approaches: The geometric surface representations
of the volumes of interest generated in step 2 are imported into a
commercial mesher. For straightforward geometric surfaces akin to those
typically obtained from CAD designs most commercial meshers provide good
tools. However for problems more typically seen in applications from 3D
imaging meshing approach in adopted commercial packages suffer a number
of drawbacks
PROBLEMS
v) Difficulty in meshing complex volumes successfully
except with very small element base length sizes
vi) where two or more structures are in contact typically you
will not get a properly conforming interface (gaps/overlaps)
vii) The connection between the mesh and the original greyscale
data that spawned it is lost - this connection can be useful for
assigning inhomogeneous material properties throughout a structure (say
bone) based on the signal strength in the parent 3D image (say
Hounsfield number to density/Young's modulus) Admittedly as a way around
this one can re-reference a posteriori the 3D data to assign material
properties to mesh elements although this requires the use of yet
another non-integrated software tool.
"Voxel based" approaches
Step 2 "Voxel based" approaches: "Voxel based" approaches convert the
segmented data (masks) directly into finite element meshes bypassing the
conversion step to geometric surface description.
By bypassing the CAD surface generation step the problems ((i) to (vii))
listed above are avoided and a very robust and automated approach can be
implemented. However early implementations did suffer from a number of
distinct drawbacks including (1) the generation of unsmooth interfaces
(originally voxels were simply converted into brick elements leading to
'lego' like models with stepped boundaries (2) Lack of adaptive meshing
- models consisted of elements of the same base length - the mesh
density was therefore constant throughout volume. These problems have
been successfully addressed by our academic group and the algorithms
developed have been implemented into a commercially available software
suite ScanFE.
The techniques developed provide, I believe unarguably, both
quantitatively and qualitatively better meshes from 3D data than
previously possible for a large class of problems.
The meshing techniques developed:
(1) can be applied to masks/volumes of interest of arbitrary
complexity
(2) To any number of masks simultaneously
(3) Generate smooth interfaces
(4) The process is automated, fast (typically 5 minutes on a pc) and
robust (guaranteed high element qualities)
In addition
(5) Topological/morphological accuracy of models only limited by
imaging accuracy - in other words model is as faithful as you can get
from the image.
(6) Conforming at contact surfaces/interfaces. (where two or more
masks are in contact in the segmented image the meshes generated will be
in contact with node to node correspondence across interface-no gaps or
overlaps)
(7) Adaptive meshing techniques are integrated within the meshing
(8) Material properties can be assigned throughout any of the meshed
masks based on parent voxel signal strength
(9) An RP model can be generated which is an exact replica of the FE
model (useful for experimental corroboration)
(10)Any ad hoc modifications to segmentation can be reflected
straightforwardly and automatically in the mesh as the process from
image segmentation to mesh generation is integrated.
The meshing algorithms to date provide either pure tet meshes or mixed
hex tet meshes but not pure hex meshes. (Both linear and mid-side noded
elements can be generated).
We have run a number of case studies which demonstrate the versatility
and robustness of the approach including:
(1) Mesh generation of a foam. Very convoluted shape (mask) was
meshed in minutes and a large deformation analysis was carried out in
Abaqus on a pc. Note shown deformation in avi is unscaled (true
deformation-courtesy www.firstnumerics.com
). We have also modeled air flow
through foam using FLUENT (CFD). (Both interstitial spaces and foam cell
walls can be meshed simultaneously for fluid-structure interaction
problems).
( http://www.simpleware.com/applications/casestudies/foam.php )
(2) Generation of a hip model based on in vivo CT data. This is a
model which includes a number of parts and a contact surface at cup
-implant head interface all generated within ScanFE and solved on a pc.
( http://www.simpleware.com/applications/casestudies/hip.php )
(3) Mesh generation of Beetle mandible - this is an impressive model
generated literally in minutes based on micro-ct data provided by Dr
Thomas Hornschmeyer. (
http://www.simpleware.com/applications/casestudies/beetle.php )
I hope this is a useful overview of possible routes to mesh generation
from 3D data. I have deliberately omitted template meshing as it is not
speaking a generic meshing technique. (Template meshing can be useful
where a large number of very similar structures need to be meshed - a
template/reference mesh is generated manually and then scaled/morphed
usually based on landmarks in image to provide patient specific versions
of the template mesh)
Philippe
Dr. Philippe G. Young
Senior Lecturer
School of Engineering and Computer Science
University of Exeter
Harrison Building
North Park Road
Exeter, EX4 4QF, UK
Phone: +44 (0)1392 263684
Fax: +44 (0)1392 263620
also
Simpleware Ltd.
Innovation Centre, University of Exeter
Rennes Drive, Exeter, EX4 4RN, UK
Website: www.simpleware.com
-----------------------------------------------------------------
To unsubscribe send SIGNOFF BIOMCH-L to LISTSERV@nic.surfnet.nl
For information and archives: http://isb.ri.ccf.org/biomch-l
Please consider posting your message to the Biomch-L Web-based
Discussion Forum: http://movement-analysis.com/biomch_l
-----------------------------------------------------------------
imaging data which I hope will be useful.
I must state that I am one of the founding members of Simpleware Ltd
which markets the software package ScanFE for the generation of finite
element (and CFD) meshes from 3D image data - however I am also an
academic with a long standing interest in this area and I would like to
think that the comments below provide, as far as possible, an impartial
assessment of the different routes available to converting CT data to FE
mesh.
There are essentially two distinct approaches to the generation of
finite element meshes from 3D imaging data (as obtained from MRI, CT,
micro-CT for example): CAD based and "voxel" based approaches.
Both approaches require the segmentation of the 3D data; in other words
the identification of volumes of interest within the image.
Step 1: Segmentation
Segmentation techniques can range from simply defining a threshold (e.g.
anything above a certain greyscale is bone) to sophisticated
combinations of complex image processing algorithms. Segmented images
consist of one or more volumes of interest (flagged voxels within the 3D
image representing bone, ligaments, muscle, fat..) which are often
called masks. Segmentation tools are provided by a wide range of
software packages including, in addition to those listed in previous
emails, Analyze (from Mayo clinic) and ScanFE.
Once masks/volumes of interest have been obtained using segmentation
tools two distinct routes to generating meshes are possible.
"CAD based" approaches
Step 2 for "CAD based" approaches: convert the segmented masks (volumes
of interest) into geometric surface representations bounding the volumes
of interest: examples of such representations include tessellated
surfaces (described by primitives such as triangles), stacks of contours
which then need to be lofted into surfaces, or NURBS type patches
(higher order polynomial representations) of the boundaries. The
conversion from volume data to surface representation is often fraught
with difficulties:
PROBLEMS
i) gaps frequently appear in surface description which
need to be painstakingly corrected manually,
ii) where two or more volumes are in contact their surface
description are usually non-conforming (gaps/overlaps)- this is a
significant problem for modeling interfaces particularly contact
surfaces,
iii) stacking two dimensional contours (essentially a so-called
21/2 D approach) is a particularly poor approach as it cannot
satisfactorily/automatically handle bifurcations
iv) the generation of NURBS surfaces (and equally the contour
stacking approach) almost inevitably engenders approximations and the
loss of features.
Step 3 for "CAD based" approaches: The geometric surface representations
of the volumes of interest generated in step 2 are imported into a
commercial mesher. For straightforward geometric surfaces akin to those
typically obtained from CAD designs most commercial meshers provide good
tools. However for problems more typically seen in applications from 3D
imaging meshing approach in adopted commercial packages suffer a number
of drawbacks
PROBLEMS
v) Difficulty in meshing complex volumes successfully
except with very small element base length sizes
vi) where two or more structures are in contact typically you
will not get a properly conforming interface (gaps/overlaps)
vii) The connection between the mesh and the original greyscale
data that spawned it is lost - this connection can be useful for
assigning inhomogeneous material properties throughout a structure (say
bone) based on the signal strength in the parent 3D image (say
Hounsfield number to density/Young's modulus) Admittedly as a way around
this one can re-reference a posteriori the 3D data to assign material
properties to mesh elements although this requires the use of yet
another non-integrated software tool.
"Voxel based" approaches
Step 2 "Voxel based" approaches: "Voxel based" approaches convert the
segmented data (masks) directly into finite element meshes bypassing the
conversion step to geometric surface description.
By bypassing the CAD surface generation step the problems ((i) to (vii))
listed above are avoided and a very robust and automated approach can be
implemented. However early implementations did suffer from a number of
distinct drawbacks including (1) the generation of unsmooth interfaces
(originally voxels were simply converted into brick elements leading to
'lego' like models with stepped boundaries (2) Lack of adaptive meshing
- models consisted of elements of the same base length - the mesh
density was therefore constant throughout volume. These problems have
been successfully addressed by our academic group and the algorithms
developed have been implemented into a commercially available software
suite ScanFE.
The techniques developed provide, I believe unarguably, both
quantitatively and qualitatively better meshes from 3D data than
previously possible for a large class of problems.
The meshing techniques developed:
(1) can be applied to masks/volumes of interest of arbitrary
complexity
(2) To any number of masks simultaneously
(3) Generate smooth interfaces
(4) The process is automated, fast (typically 5 minutes on a pc) and
robust (guaranteed high element qualities)
In addition
(5) Topological/morphological accuracy of models only limited by
imaging accuracy - in other words model is as faithful as you can get
from the image.
(6) Conforming at contact surfaces/interfaces. (where two or more
masks are in contact in the segmented image the meshes generated will be
in contact with node to node correspondence across interface-no gaps or
overlaps)
(7) Adaptive meshing techniques are integrated within the meshing
(8) Material properties can be assigned throughout any of the meshed
masks based on parent voxel signal strength
(9) An RP model can be generated which is an exact replica of the FE
model (useful for experimental corroboration)
(10)Any ad hoc modifications to segmentation can be reflected
straightforwardly and automatically in the mesh as the process from
image segmentation to mesh generation is integrated.
The meshing algorithms to date provide either pure tet meshes or mixed
hex tet meshes but not pure hex meshes. (Both linear and mid-side noded
elements can be generated).
We have run a number of case studies which demonstrate the versatility
and robustness of the approach including:
(1) Mesh generation of a foam. Very convoluted shape (mask) was
meshed in minutes and a large deformation analysis was carried out in
Abaqus on a pc. Note shown deformation in avi is unscaled (true
deformation-courtesy www.firstnumerics.com
). We have also modeled air flow
through foam using FLUENT (CFD). (Both interstitial spaces and foam cell
walls can be meshed simultaneously for fluid-structure interaction
problems).
( http://www.simpleware.com/applications/casestudies/foam.php )
(2) Generation of a hip model based on in vivo CT data. This is a
model which includes a number of parts and a contact surface at cup
-implant head interface all generated within ScanFE and solved on a pc.
( http://www.simpleware.com/applications/casestudies/hip.php )
(3) Mesh generation of Beetle mandible - this is an impressive model
generated literally in minutes based on micro-ct data provided by Dr
Thomas Hornschmeyer. (
http://www.simpleware.com/applications/casestudies/beetle.php )
I hope this is a useful overview of possible routes to mesh generation
from 3D data. I have deliberately omitted template meshing as it is not
speaking a generic meshing technique. (Template meshing can be useful
where a large number of very similar structures need to be meshed - a
template/reference mesh is generated manually and then scaled/morphed
usually based on landmarks in image to provide patient specific versions
of the template mesh)
Philippe
Dr. Philippe G. Young
Senior Lecturer
School of Engineering and Computer Science
University of Exeter
Harrison Building
North Park Road
Exeter, EX4 4QF, UK
Phone: +44 (0)1392 263684
Fax: +44 (0)1392 263620
also
Simpleware Ltd.
Innovation Centre, University of Exeter
Rennes Drive, Exeter, EX4 4RN, UK
Website: www.simpleware.com
-----------------------------------------------------------------
To unsubscribe send SIGNOFF BIOMCH-L to LISTSERV@nic.surfnet.nl
For information and archives: http://isb.ri.ccf.org/biomch-l
Please consider posting your message to the Biomch-L Web-based
Discussion Forum: http://movement-analysis.com/biomch_l
-----------------------------------------------------------------