Multi-objective Optimization and Stochastic Analysis
in Focused Ultrasonic Therapy Simulation
T. Clees, N. Hornung, I. Nikitin, L. Nikitina and D. Steffes-lai
Fraunhofer Institute for Algorithms and Scientific Computing, Sankt Augustin, Germany
Keywords: Stochastic Modeling and Simulation, Multi-objective Optimization, Healthcare.
Abstract: We present new results in stochastic multi-objective optimization applied to focused ultrasonic therapy
planning. This type of non-invasive therapy uses focused ultrasound for the destruction of tumor cells and
magnetic resonance tomography for identification of tumor volume and healthy organs. During the therapy
planning the treatment parameters, such as frequency and intensity of ultrasound, are adjusted to achieve
maximal tumor destruction and minimal influence to the healthy organs. For this purpose multi-objective
optimization is used. RBF metamodeling is employed for continuous representation of discretely sampled
results of numerical simulation and for evaluation of inherent uncertainties. We apply two algorithms for
multi-objective optimization capable of non-convex Pareto front detection in the considered problem. Cross-
validation procedure and sensitivity analysis are used for estimation of uncertainties. A realistic application
case demonstrates the efficiency of the approach.
1 INTRODUCTION
Due to the non-invasive nature of the focused
ultrasonic therapy its control is often limited to
imaging methods, e.g. MRT. Numerical simulation
becomes an important step for the therapy planning.
Efficient methods for the focused ultrasonic
simulation have been presented in paper (Georgii et
al., 2011). It uses a combination of Rayleigh-
Sommerfeld integral for near field and angular
spectrum method for far field computations, which
allows determining the pressure field in
heterogeneous tissue. The bioheat transfer equation
is used to determine the temperature increase in
therapy region. Thermal dose is defined according to
CEM model or Arrhenius model (Nandlall et al.,
2009); (Pearce, 2009) as a functional of temperature-
time dependence in every spatial point in therapy
region. The simulation is considerably accelerated
by GPU based parallelization.
The purpose of therapy planning is a
maximization of thermal dose inside the target zone
(TDin) and minimization of thermal dose outside
(TDout). As usual in multi-objective optimization
(Ehrgott AND Gandibleux, 2002), the optimum is
not an isolated point but a hypersurface (Pareto
front) composed of points satisfying a tradeoff
property, i.e. none of the criteria can be improved
without simultaneous degradation of at least one
other criterion. Thus, for a two-objective problem,
the Pareto front is a curve on the plot (TDin, TDout)
bounding the region of possible solutions, see Fig.2.
Efficient methods have been developed for
determining the Pareto front.
The simplest way is to convert multi-objective
optimization to single objective one, by linearly
combining all objectives into a single target function
t(x)=
w
i
f
i
(x) with user-defined constant weights
w
i
. Maximization of the target function gives one
point on Pareto front, while varying the weights
allows to cover the whole Pareto front. In this way
only convex Pareto fronts can be detected, because
non-convex Pareto fronts produce not maxima but
saddle points of the target function. There are
methods applicable also for non-convex Pareto
fronts.
Non-dominated set algorithm (NDSA) finds a
discrete analogue of Pareto front in a finite set of
points. For two points f and g in optimization criteria
space the first one is said to be dominated by the
second one if f
i
g
i
holds for all i=1..Ncrit. A point f
belongs to non-dominated set if there does not exist
another point g dominating f. (Kung et al., 1975)
implements a fast recursive procedure to find all
non-dominated points in a given finite set.
43
Clees T., Hornung N., Nikitin I., Nikitina L. and Steffes-lai D..
Multi-objective Optimization and Stochastic Analysis in Focused Ultrasonic Therapy Simulation.
DOI: 10.5220/0004421900430048
In Proceedings of the 3rd International Conference on Simulation and Modeling Methodologies, Technologies and Applications (SIMULTECH-2013),
pages 43-48
ISBN: 978-989-8565-69-3
Copyright
c
2013 SCITEPRESS (Science and Technology Publications, Lda.)
Figure 1: MRT slices are used as a basis for volumetric material model.
Figure 2: Robust multi-objective optimization using
metamodeling of simulation results. Red points denote
simulation results, the continuous green region indicates a
metamodel of simulation results, the blue curve the Pareto
front, the black point with error bars an optimal
representative with confidence limits.
A continuous method is LP-based local
improvement algorithm (LIA). It produces a
trajectory towards the Pareto front starting from an
initial design. Each step solves a linear program
(Hornung et al., 2010)
maximize
;
where (f
i
,x) 0 and - x
j
w.r.t. the threshold
and the step x, for the given
gradients of objectives f
i
and the size of trust
region [-,]. Here i=1..Ncrit, j=1..Npar. The
approach ensures that all criteria have an
improvement at least
, maximally possible within
the given trust region. The algorithm terminates at
Pareto front, where no further improvements are
possible.
In the case when only a restricted number of
simulations is available metamodeling of simulation
results becomes useful. The objectives are
interpolated continuously in between simulation
results, and optimization algorithms are applied to
interpolated functions. In particular, RBF
metamodeling (
Buhmann, 2003) has the advantage
of generic non-degeneracy in arbitrary dimensions
and the availability of a tolerance predictor. The
interpolated function f(x) is represented as a linear
combination of special functions () depending only
on the distance to the sample points x
k
:
f(x) =
k=1..Nex
p
c
k
(|x-x
k
|)
(1)
The coefficients c
k
can be found from known
function values in sample points f(x
k
) by solving a
moderately sized linear system with a matrix
kn
=(|x
k
-x
n
|).
In practical problems, different sources of
uncertainty should be considered. Uncertainties of
metamodeling are related with the quality of
sampling in parameter space. These uncertainties are
reduced with the increasing density of sampling, i.e.
SIMULTECH2013-3rdInternationalConferenceonSimulationandModelingMethodologies,Technologiesand
Applications
44
when more simulations are included into analysis.
This type of uncertainty can be estimated by a cross-
validation procedure, measuring a sensitivity of the
function value to the removal of sample points. With
the usage of Sherman-Morrison-Woodbury formula
it can be found analytically (Nikitin et al, 2012):
k
= f(x
k
) - f(x
k
)|
xk remove
d
= c
k
/(
-1
)
k
k
(2)
Another type of uncertainties comes from physical
model. In addition to optimization parameters x, the
objectives also depend on a number of model
parameters p, such as physical properties of
biomaterials. These parameters are measured
imprecisely. The corresponding variation of results
can be estimated by means of the first order
reliability method FORM (Clees et al., 2012):
(f) = (
j
=1..N
p
a
r
(f/
p
j
(p
j
))
2
)
1/2
(3)
where non-correlated variations of parameters p
j
are
assumed, and () denotes standard deviation.
Finally, these two sources of scatter can be
combined into a single measure:
err = (
2
+
2
)
1/2
.
(4)
In further sections we apply the above described
methodology for the multi-objective optimization
and stochastic analysis of focused ultrasonic therapy
simulation.
2 MODELING
A generic workflow for modeling of the focused
ultrasonic therapy has been described in our paper
(Borsotto et al., 2012). Numerical simulation uses
FUSimlib software (http://www.simfus.de ) on 512 x
512 x 256 voxel grid. Ultrasound has been focused
in the center of the target zone for the neutral breath
state. The computational model includes angular
spectrum method with heterogeneous tissue and
reflections. CEM model is used for determination of
thermal dose. The result after 10 seconds of
exposure time has a form of spatial distributions of
pressure amplitude, temperature and thermal dose.
Fig.3 shows a typical result for thermal dose on slice
97/256 near the focal point.
The registration of tissue deformation on the
basis of MRT images is used for proper
characterization of the breathing process. It has been
shown in (Borsotto et al., 2012) that the breathing
process is one of the major sources of uncertainties
in focused ultrasonic therapy simulation. For
characterizing the breathing process two image
sequences have been processed, each containing 104
z-slices with 320 x 320 xy-resolution. One sequence
corresponds to the breath-in state, another one to the
neutral breath state. Point-to-point correspondence
between images has been determined using a block-
matching method. Fig.1 shows the resulting field of
motion vectors for slice 51/104 of the neutral breath
sequence.
Figure 3: Typical simulation result, thermal dose (range:
0eq.min/blue to 0.6eq.min/red). The target zone in the
neutral breath state is marked by the white circle.
Figure 4: Spatial material distribution: gel (black), liver
(green), other soft tissue (red), cartilage (blue), bones
(yellow). White color marks the target zone.
Segmentation procedure defines spatial distribution
of biomechanical characteristics on the voxelized
model. 5 types of materials are introduced shown
with different colors on Fig.4, for the slice 21/104. A
target zone has been modeled as a ball of 1cm
diameter located on the upper part of the liver. Its
motion according to breathing process is evaluated
on the basis of the motion vector field found in the
registration procedure and controlled by a single
parameter (t=0…1) for the transition from neutral to
breath-in state. A finer 512 x 512 x 256 voxel grid
Multi-objectiveOptimizationandStochasticAnalysisinFocusedUltrasonicTherapySimulation
45
has been used for simulation, where the distribution
of material has been subsampled into, using proper
positioning transformations.
3 OPTIMIZATION
Optimization is performed with respect to the
following parameters.
Table 1: Optimization parameters and their variations.
frequency 0.25…0.75 MHz
initial particle speed 0.23...0.282 m/s
The frequency of transducer is an important
parameter controlling focused ultrasonic therapy
simulation. The other one, initial particle speed, is
proportional to an acoustic intensity emitted by the
transducer (Georgii et al., 2011). As optimization
objectives the thermal dose inside and outside the
target zone have been defined as sums of the thermal
dose over corresponding voxels, TDin / TDout.
The variation range of optimization parameters was
regularly sampled with 25 simulations, from which
19 fall in the region of interest, shown on Fig.5. RBF
metamodel constructed on simulation results shows
Pareto front of non-convex type, for which NDSA
and LIA Pareto front detectors have been applied.
Figure 5: Results of multi-objective optimization. Red
points denote simulation results, the green points indicate
a metamodel of simulation results, the blue points indicate
the Pareto front detected by NDSA, the magenta points
show trajectories of interior points towards Pareto front
found with LIA. Black crosses show the tolerances of the
metamodel found with cross-validation procedure.
For convenience of the therapy planning it is
advisable to use a special tool for design parameter
optimization, DesParO (Clees et al., 2012), see
Fig.6. It allows to change interactively optimization
parameters and to see immediately the variation of
optimization criteria. Constraints can be set e.g.
maximizing one objective and minimizing the other,
in this way the Pareto front can be explored.
Graphical representation of interdependencies
between parameters and criteria allows to find most
influencing parameters and most sensitive criteria.
Also, the uncertainties of metamodeling found with
cross-validation procedure are shown (the red bars
under criteria sliders).
Further we focus on uncertainties coming from
physical model.
Figure 6: Optimization problem in DesParO Metamodel
Explorer. On the top: sliders of parameters, in the center:
sliders of criteria, on the bottom: a pattern of
interdependencies between parameters and criteria.
4 SENSITIVITY ANALYSIS
In our previous paper (Borsotto et al, 2012) the
sensitivity of the result to variation of 29 parameters
has been evaluated, including 5 biomechanical
SIMULTECH2013-3rdInternationalConferenceonSimulationandModelingMethodologies,Technologiesand
Applications
46
characteristics for 5 materials each, 3 blood
characteristics and 1 breathing parameter. It has
been shown that the parameters most influencing the
result are the absorption coefficients for gel, soft
tissue and liver as well as the breathing parameter.
For improving accuracy of the physical model a
precise measurement of absorption coefficients as
well as synchronization of transducer focal point
with breathing process has been proposed.
Here we present a new estimation of sensitivity,
based on improved measurements of critical
parameters given in (Peters, 2007).
For estimation of sensitivity, the parameters have
been varied in ± interval according to Table 2
below, then a central difference scheme has been
used to estimate entries of the Jacobian matrix:
J
i
j
= crit
i
/
par
(5)
The sensitivity matrix is defined as
S
i
j
= J
i
j
j
(6)
and presented in Table 3. Then total r.m.s. of two
criteria is evaluated:
(crit
i
) = (
j
=1..N
p
a
r
S
i
j
2
)
1/2
(7)
and presented in Table 4 together with the mean
values of the criteria.
Comparing with the previous result, we see
significant improvement of precision, especially if
breathing influence is compensated. These
uncertainties should be taken into account during the
therapy planning.
Table 2: Standard deviations of absorption coefficient (in
1/m), from (Peters, 2007).
gel
0.001
liver
1.2
soft tissue
0.7
Table 3: Sensitivity matrix for objectives TDin / TDout
(in eq.min).
gel soft
tissue
liver
absorption -0.2
/ -0.3
-183.6 /
277.3
709 /
378.5
breath -505.2 / 546.6
Table 4: Standard deviations and mean values for TDin /
TDout (in eq.min).
TDin TDout
r.m.s.
total
889.5 720.4
r.m.s.
breath
compens.
732.1 469.2
mean 2273.7 3743.8
5 CONCLUSIONS
A generic approach for focused ultrasonic therapy
planning on the basis of numerical simulation has
been presented including multi-objective
optimization and stochastic analysis. Its application
to a realistic test case has been demonstrated. RBF
metamodeling of simulation results has been
performed for continuous representation of two
optimization objectives. Non-convex Pareto front of
the objectives has been determined by means of non-
dominated set and local improvement algorithms.
Uncertainties of metamodeling have been estimated
by means of cross-validation procedure. These
uncertainties can be reduced with the increasing
density of sampling, i.e. including more simulations
into analysis. Uncertainties of physical model have
been estimated by means of sensitivity analysis.
Improved accuracy of the physical model and
compensation of the influence of the breathing
process provide better precision of the result.
ACKNOWLEDGEMENTS
We are grateful to our colleagues from Fraunhofer
MEVIS: Dr. Caroline v. Dresky, Dr. Sebastian
Meier, Dr. Joachim Georgii for providing us with
MRT images, a segmented model, simulation
software as well as for fruitful discussions. This
work was supported by the Fraunhofer Internal
Program under Grant No. MAVO 821012 (FUS).
REFERENCES
Georgii, J., Dresky, C. V., Meier, S., Demedts, D.,
Schumann, C., Preusser, T., Focused Ultrasound -
Efficient GPU Simulation Methods for Therapy
Planning, in Proc. Workshop on Virtual Reality
Interaction and Physical Simulation VRIPHYS, Lyon,
France, 2011, J. Bender, K. Erleben, and E. Galin
Multi-objectiveOptimizationandStochasticAnalysisinFocusedUltrasonicTherapySimulation
47
(Editors), Eurographics Association 2011, pp. 119-
128.
Nandlall, S., Arora, M., Schiffter, H. A., Coussios, C.-C.:
On the Applicability of the Thermal Dose Cumulative
Equivalent Minutes Metric to the Denaturation of
Bovine Serum Albumin in a Polyacrylamide Tissue
Phantom. Proc. 8th Int. Symp. Therapeutic Ultrasound
(AIP), 1113:205-209, 2009.
Pearce, J. A.: Relationships between Arrhenius models of
thermal dose damage and the CEM 43 thermal dose.
Energy-based Treatment of Tissue and Assessment V
(Proceedings of SPIE), editor: Ryan, T.P., 7181, 2009.
Ehrgott, M., Gandibleux, X., (Eds.), Multiple criteria
optimization: state of the art annotated bibliographic
surveys, Kluwer 2002.
H. T. Kung, F. Luccio, and F. P. Preparata. On finding the
maxima of a set of vectors. Journal of the ACM,
22(4):469–476, 1975.
Hornung, N., Nikitina, L., Clees, T., Multi-objective
optimization using surrogate functions, in Proc. 2nd
International Conference on Engineering
Optimization, September 6-9, 2010, Lisbon, Portugal.
Buhmann, M. D., Radial Basis Functions: theory and
implementations, Cambridge University Press, 2003.
Borsotto D., Clees T., Nikitin, I., Nikitina, L., Steffes-lai
D., Thole C.-A., Sensitivity and robustness aspects in
focused ultrasonic therapy simulation, in J.Herskovits,
Ed., CDROM Proc. EngOpt 2012, third Int. Conf. on
Engineering Optimization, Rio de Janeiro, Brazil, July
1-5, 2012, #268.
Nikitin, I., Nikitina, L., Clees, T., Nonlinear metamodeling
of bulky data and applications in automotive design, in
M. Günther et al (eds), Progress in industrial
mathematics at ECMI 2010, Mathematics in Industry
(17), Springer, 2012, pp. 295-301.
Clees, T., Nikitin, I., Nikitina, L., Thole, C.-A., Analysis
of bulky crash simulation results: deterministic and
stochastic aspects, in N.Pina et al (Eds.): Simulation
and Modeling Methodologies, Technologies and
Applications, AISC 197, Springer 2012, pp.225-237.
Peters, K., Experimentelle Untersuchungen zur
nichtinvasiven Gewebeablation durch
hochenergetischen fokussierten Ultraschall (HIFU),
Dissertation 2007, Ludwig-Maximilians-University,
Munich (in German).
SIMULTECH2013-3rdInternationalConferenceonSimulationandModelingMethodologies,Technologiesand
Applications
48