Karol Miller
Challenges for computationalbiomechanics for medicine
Karol MillerVisiting Professor. University of Luxembourg
Intelligent Systems for Medicine Lab.
The University of Western Australia35 Stirling HighwayCrawley WA 6009, AUSTRALIAEmail: [email protected]://www.mech.uwa.edu.au/~kmillerhttp://school.mech.uwa.edu.au/ISML/
Institute of Mechanics and Advanced Materials
Karol Miller
Perth
The University of Western Australia
Karol Miller
Russell Taylor’s prophecy:
The market for scientific computations in medicine would be as large as in engineering by 2020
Computer-Integrated Surgery (CIS) systems will improve clinical outcomes and the efficiency of health care delivery. CIS systems will have a similar impact on surgery to that long since realised in Computer-Aided Design (CAD) and Computer-Integrated Manufacturing (CIM).
Karol Miller
Oden, Belytschko, Babuska, Hughes:
One of the greatest challenges for mechanists is to extend the success of computational mechanics to fields outside traditional engineering, in particular to biology, biomedical sciences, and medicine
Karol Miller
GENG4405
5
Karol Miller
Our main motivation - image-guided neurosurgery Image of brain tumour (green)
is superimposed on patient as an aid to surgical planning and
navigation
Courtesy of SPL, Harvard
Karol Miller
Courtesy Prof. Wies Nowinski, A-Star, Singapore
The brain is complicated…
But we only wish to compute displacements
Karol Miller
Gargantuan challenges:
1. For biomechanical computations to be practical in a clinical environment, computational grids must be obtained from standard diagnostic medical images automatically and rapidly.
2. Real-time computations on commodity hardware
3. Real-time simulation of cutting, damage and propagation of discontinuities
4. Mathematical formulations that are weakly sensitive to uncertainties in mechanical properties of tissues are necessary.
Karol Miller
Gargantuan challenges:
1. For biomechanical computations to be practical in a clinical environment, computational grids must be obtained from standard diagnostic medical images automatically and rapidly.
The current practice of patient-specific model generation involves image segmentation and finite element meshing. Both present themselves as formidable problems that are very difficult to automate. Entirely novel approaches are needed.
Karol MillerJoldes et al. (2009), MICCAI 2009, Part II, LNCS 5762, pp. 300-307
Many days of tedious work
Challenge 1: efficient generation of patient-specific computational grids from medical images
Karol Miller
c)
d) e)
Patient-Specific Finite Element MeshesJoldes et al. (2009), MICCAI 2009, Part II, LNCS 5762, pp. 300-307
Karol Miller
Neuroimage as a computational model?
2D MRI slice “Hard” segmentation Assignment of mechanical properties based on statistical tissue classification
0
3000
[Pa] 6000
Tumour
Brain
Ventricle
Karol Miller
Comparison between FE model and fuzzy mesh-free model constructed respectively from segmentation and fuzzy tissue classification. (a) T2 MRI of the brain with the tumour and ventricles present, notice that no clear boundaries can be easily defined, especially for the tumour, (b) finite element model of ventricles generated from segmentation, (c) finite element model of the tumour generated from segmentation, (d) fuzzy tissue classification of ventricle, (e) fuzzy tissue classification of tumour,
(f) fuzzy mesh-free model of ventricle and (g) fuzzy mesh-free model of tumour, green dots represent nodes while grey grids represent uniform background integration grids. Notice that nospecific tissue class is defined in the domain. Material properties are assigned directly to theintegration points based on fuzzy classification results.
Zhang et al. (2013), IJNMBE 29(2), pp. 293–308
Karol Miller
3D patient-specific meshless computational grid of the brain
Miller et al. (2012), J. Biomech. 45(15), pp. 2698-2701
Green – parenchyma
Red – ventricles
Blue - tumour
Karol Miller
Evaluation of accuracy for three cases
Left column: Finite Element Models, with parenchyma, tumour (red) and ventricle (blue) modelled separately.
Middle column: Fuzzy Mesh-free Model without explicitly separating the tumour and ventricles, fuzzy tissue classifications of tumour (red), and ventricle (blue) are shown as cloud superimposed on the image; Nodes are shown asgreen dots.
Right column: Difference of the simulation results (computed deformation field) from the two models over the whole problem domain [mm].
Karol Miller
The 'double doughnut' General Electric 1.5T open magnet at the Brigham and Women's Hospital, Boston seen end-on (left) and from the side (right), recently replaced by AMIGO
Karol Miller
Whole-body meshless model for CT registration
source image target image
Whole-body meshless model. Tissue properties are assigned automatically to integration points, based on fuzzy classification
Li et. al (2014) Medical Image Analysis
Displacements of the order of 10 cm
Karol Miller
Evaluation of registration accuracy
Li et. al (2014) Medical Image Analysis
The dotted line and dashed line (they are nearly overlapping) represent lung contours extracted fromimages registered using deformations predicted by means of the meshless model used in this studyand previously validated finite element model. The solid line is the lung contour extracted from the
target image.
Karol Miller
Gargantuan challenges:
2. In surgical simulation interactive (haptic) rates (i.e. at least 500 Hz) are necessary for force and tactile feedback delivery. In intra-operative image registration one needs to provide a surgeon with updated images in less than 40 seconds.
To achieve these, real-time computational speeds for highly non-linear models with at least 100,000 degrees of freedom must be achieved on commodity computing hardware.
Joldes et al. (2010) Computer methods in applied mechanics and engineering 199 (49), 3305-3314
Karol Miller
Graphics Processing Unithttp://www.gpucomputing.net/ Computational Biomechanics Community http://gpucomputing.net/?q=node/218
1536 cores
Karol Miller
What does GPU like?• Problems that can be expressed as data-parallel computations –
the same program is executed on many data elements in parallel
What does GPU not like?•Communications (between cores and especially with CPU and
external devices)
Explicit algorithms are therefore preferable
Karol Miller
Deformation No. of elements
Computation time (s) GPU Speed up (x)Abaqusstatic
CPU GPU Abaqus static
CPU
Compression
48000
3732 57.7 1.76 2120 32.7
Extension 1087 69.1 2.37 458.6 29.1
Amazing performance!Comparison of computational times when using GPU and CPU
Karol Miller
Mesh Numberofnodes
Number of elements SkullHexa Linear
tetrasANPtetras
Totalelements
Numberof nodes
Numberoftriangles
Original 12693 10596 4831 1398 16825 1993 3960
Refined 95669 84768 32439 8085 125292 7945 15840
Mesh No. of steps required for convergence (δ = 10E-4)
Run time for 3000 steps (s)
Speed up (x)
CPU GPU CPU GPU
Original 1887 2103 79.7 3.54 22.5
Refined 3120 3091 543.4 19.95 27.2
Table 1: Structure of the brain meshes used
Table 2: Computation times for brain shift simulation
Wittek et al. (2010) Progress in Biophysics and Molecular Biology, 103, 292-303
In comparison Courtecuisse et al. (2014) Medical Image Analysis 18 394–410
has a brain model with 1734 nodes…
Karol Miller
Head impact simulation (time-accurate)
Computations conducted on a PC with a Tesla C1060 GPU having 4 GB of RAM and 240 cores.
Karol Miller
• nodes: 1101559• elements: 1061799 - hexas• Computation time: 40000 steps in 15 minutes
Karol Miller
Using our algorithms on GPU’s can potentially allow computing large, non-linear problems between 500 and 5000 times faster than using commercial software on standard computers.
Close-to-real-time interactive use of FEM computations for design seems to be within reach. And this can be achieved on computing hardware costing ca. $10000!
General large nonlinear engineering computations that are currently most often subcontracted to specialized consultancies will be possible on desktop computers (such as Tesla Supercomputer).
Design engineers will be able to run simulations of their design concepts interactively, greatly increasing the number of cases they are able to consider.
Broader impact on the practice of engineering computations
Karol Miller
Gargantuan challenges:
3. Surgical manipulation involves not only large deformations of soft tissues but also cutting and (often unintentional) damage.
Modelling and real-time simulation of cutting, damage and propagation of discontinuities remains an unsolved and very challenging problem of computational biomechanics.
But some progress reported in Courtecuisse et al. (2014) Medical Image Analysis 18 394–410Jin et al. (2014) Computer Methods in Biomechanics and Biomedical Engineering.17(7) 800-811
Karol Miller
Gargantuan challenges:
4. Human soft tissues are highly variable, and despite recent progress in magnetic resonance (MR) and ultrasound elastography, their in-vivo properties are difficult to obtain.
Therefore mathematical formulations that are weakly sensitive to uncertainties in mechanical properties of tissues are necessary.
Some progress reported in Miller and Lu (2013) Journal of the Mechanical Behavior of Biomedical Materials.
27, 154-166
Karol Miller
Fromhttp://euromech534.emse.fr/ ttp://euromech534.emse.fr/To this end, it becomes a common practice to combine video based full-field
measurements of the displacements experienced by tissue samples in vitro with a custom inverse method to infer, using nonlinear regression, the best-fit material parameters. Similar approaches also exists for characterizing tissues in vivo where advanced medical imaging can provide precise measurements of tissue deformation under different modes of action and inverse methodologies are used to derive material properties from those data.
But perhaps we can obtain useful, patient-specific results WITHOUT the knowledge of patient-specific mechanical properties of tissues?
Karol Miller
If our loading is through the enforced motion of boundary conditions (dimension [mm]) and our result is a displacement field (in [mm]), this result cannot depend on a stress parameter (dimension [Pa]).
The result may still depend on Poisson’s ratio, but not for almost incompressible materials.
This suggests that if we are able to formulate our biomechanical investigations as Dirichlet problems (i.e. problems driven by enforced motion of boundaries) we can expect to obtain meaningful patient-specific results without knowledge of patient-specific properties of tissues.
Simplistic, homogeneous linear-elastic case
Karol Miller
Extension of cylindrical samples (Miller, J. Biomech. 2001) – deformed shape does not depend on mechanical
properties
Z/H
f(Z)
Sides of deformed samples for Neo-Hookean and Extreme Mooney material models for extensions h/H=1.1, 1.2, 1.3
Analogical result for compression (Miller, J. Biomech. 2005)
Karol Miller
Center of Gravity Displacements (mm)Material Model/Analysis Type
Ventricles Tumor
∆X ∆Y ∆Z ∆X ∆Y ∆Z
MRI Determined 3.4 0.2 1.7 5.5 -0.2 1.7
Hyperviscoelastic material/Geometrically non-linearanalysis
2.6 -0.1 2.1 5.2 -0.4 2.7
Hyperelastic material/Geometrically non-linearanalysis
2.6 -0.1 2.1 5.2 -0.4 2.7
Linear elastic material/Geometrically non-linearanalysis
2.6 -0.1 2.1 5.0 -0.5 2.7
Linear elastic material/Linear analysis 0.7 0.2 1.9 3.7 -0.3 2.6
Image registration: results
32
Karol Miller
CONCLUSIONS - the challenges awaiting us:
1. For biomechanical computations to be practical in a clinical environment, computational grids must be obtained from standard diagnostic medical images automatically and rapidly -> possible solution: meshless solution methods with fuzzy tissue classification (but perhaps something like cutFEM can be better but as yet no demonstration for realistic nonlinear problems exists…)
2. Real-time computations on commodity hardware -> possible solution: use GPUs
3. Real-time simulation of cutting, damage and propagation of discontinuities -> ???
4. Mathematical formulations that are weakly sensitive to uncertainties in mechanical properties of tissues are necessary -> reformulate as Dirichlet problems?
Karol Miller
Acknowledgements:FNR, University of Luxembourg and Stephane!Prof. Ron Kikinis (Harvard) Prof. Simon Warfield (Harvard)Prof. Kiyoyuki Chinzei (AIST)Dr Toshikatsu Washio (AIST)Prof. Adam Wittek
(ISML, UWA)A/Prof. Grand Joldes
(ISML, UWA)and many very talented research students
Funding: ARC, NHMRC, THANK YOU NIH, NVIDIA,
Leverhulme Trust
Karol Miller
Look for it in the bookstore near you…
Karol Miller
And these as well…
