ZIB: Optimization and Thermoregulation

MATHEON-A1

Optimization and Thermoregulation

In the regional hyperthermia cancer therapy, radio frequency radiation is used to heat the tumor in order to make it more susceptible to an accompanying radio- or chemotherapy. Heat transport by blood flow has a major impact onto the attainable tumor temperature. In this project, a multiscale model of the heat transport in the human vascular system is developed and implemented in order to obtain more accurate simulations, and in order to find optimal antenna parameters. For the solution of the arising stationary, periodic, or transient optimization problems we consider function space oriented interior point methods. These are inexact Newton pathfollowing methods in function space. We study control reduced primal interior point methods for optimal control problems with PDEs and pointwise control or state constraints, and Lavretiev regularization. For interior point methods applied to optimal control problems with state constraints the importance of rational barrier functionals has been discovered. In the case of finite dimensional control a generically optimal rational order could be determined. A path-following method in function space has been developed that exploits problem structure by a pointwise damping step and problem suited adaptivity. The combination of Lavrentiev regularization and control reduction turned standard software (COMSOL) for the solution of non-linear PDEs into an efficient solver for optimal control problems.

Short Description

For many years heat has been studied as a therapeutic option in medicine, e.g. in the various realizations of the cancer therapy hyperthermia. Part of the members have been successfully involved in therapy planning for regional hyperthermia in close cooperation with radiologists and surgeons at the Charité Berlin. In the course of this cooperation a number of important mathematical problems have come up that could not be treated within that medicine dominated project, but, when successfully solved, would have the potential to open a new reliable perspective for the general application of heat within medical treatment.

These problems include (1) a more realistic modelling of the temperature distribution within the body and (2) an optimization of antenna parameters in the setting of refined temperature distribution models and a more adequate objective function. The solution of these problems requires both a firm analytical derivation and an efficient algorithmic realization. In this close context, (3) interior point methods in function space are also developed further for the benefit of a class of optimal control problems in PDEs.

Multiscale models of thermoregulation

The transport of heat by blood flow contributes significantly to the final temperature distribution in the human body. Up to now, the geometric structure of the vascular system has been neglected by imposing an isotropic, homogeneous Helmholtz term that accounts for the heat transport.

Since the directional transport of heat by the vascular system is considered significant, it should be taken into account. For computational tractability, a multiscale model of the vascular system has to be developed, that treats large vessels directly as 3D objects, medium size vessels as 1D simplifications, and small vessels by a homogenized Helmholtz term as before. The development of multiscale perfusion models has been pursued mainly on the scale of medium size vessels with a diameter between 10mm and 2mm. These have a non-negligible thermal influence and their geometry can be acquired individually from MR and CT scans. However, they are too small to be represented as 3D objects in a coarse mesh at reasonable computational cost. A representation of medium scale vessels as sequences of 1D edges in tetrahedral meshes and their coupling to the surrounding tissue based on conservation of mass flow has been analyzed.


Large size vessels.	Embedding of medium size vessels.

Another important effect of medium scale vessels is their influence on the tissue perfusion due to the flow balances in the vascular system. In particular, the high reduction of peripheral resistance in heated muscle may lead to decreased blood supply to a tumor fed by the same vessel. This steal-effect is observed in clinical practice. A hierarchical model of flow balances coupled to the bio-heat-transfer equation is developed.

The steal effect.

The perfusion of the tumor is highly individual, and a-priori models are unreliable. Blood flow to and from the tumor can be measured by MRI using contrast agents. We model the tumor vasculature as a porous medium, giving the resulting blood flow to be used in therapy planning.

Interior point methods for elliptic and parabolic PDE

Interior point methods have been proven to be a very efficient class of methods for inequality constrained finite dimensional optimization. We consider adaptive interior point methods for the solution of optimization problems in infinite dimensional function spaces. To obtain efficient optimization algorithms it is important to analyse and exploit the analytical structure of the problem. Once a proveably convergent method in function space is found, an inexact Newton pathfollowing algorithm can be constructed on the base of an efficient discretization scheme. The framework of inexact Newton methods yields accuracy requirements on the discretization error used to control the adaptive mesh refinement.

Path-following in function space.

Our work envolves various topics in control theory. We developed a primal interior point method for parabolic problems with linear equation, observation at the final time in the objective functional, and pointwise state constraints and proved convergence for a logarithmic barrier function. This idea gains particular importance in the context of transient and periodic heating strategies.

One highlight of our theoretical approach for problems with state constraints is a new technique to analyze them directly, without continuity assumptions for the state. This idea has the potential to extend the derivation of optimality conditions to a large class of problems.

We also investigate different regularization techniques for steady-state heating. In particular, the application of interior point methods and their coupling with a Lavrentiev type regularization have been studied. By this regularization technique, the original state constraints, e.g. $\theta_{min}\leq \theta \leq \theta_{max}$ , are replaced by $\theta_{min}\leq \theta +\varepsilon u \leq \theta_{max}$ , with a small regularization parameter $\varepsilon$ , where is the control function. Based on the analysis of the barrier homotopy path an adaptive path-following algorithm in function space was constructed and analysed. In particular, it has been shown that this algorithm produces a sequence of iterates that converges to the solution of the original state constrained problem. An implementation for our hyperthermia treatment planning problem shows that this regularization improves the performance of our algorithms significantly.

Moreover, we consider control reduced primal interior point methods where in the case of bound constrained control, the control can be eliminated from the optimality system. Then the reduced system only contains smooth variables. On the one hand this leads to optimal error estimates, on the other hand the analytical properties of the underlying differential equation can be exploited more efficiently, and we obtain a superlinearly convergent algorithm. Besides, pathfollowing can be performed on relatively coarse discretizations, which makes the resulting algorithm very efficient.

A further topic of interest is the minimization of an $L^{\infty}$ -functional constrained by an elliptic PDE. After transformation into a problem with a linear objective functional and additional state constraints, this problem can be handled without additional regularization.

Comparison of $L^{2}$ (left) and maximum-norm objective (right) on a stationary problem with discontinuous coefficients.
The circular target area is covered more evenly by the solution of the $L^{\infty}$ objective.

State constrained problems and their regularization

State constraints are an important and difficult aspect of optimal control with PDEs. Analytically, they are well understood in the case of continuous states, while steps towards an understanding in the remaining cases have only been taken recently. Their high non-linearity necessitates the use of regularization and path-following methods for their solution. We focused on two regularization approaches and their interplay. For problems with continuous states, interior point techniques can be applied. Here the state constraints are replaced by a logarithmic or rational barrier term, and the original problem is tackled by solving a sequence of non-linear problems. For cases, where discontinuous states cannot be excluded in advance, these techniques have been complemented by Lavrentiev type regularization. Here the state constrained problem is transformed into a problem with mixed control-state constraints, which is more regular and easier to solve.

Interior point methods for stationary problems with state constraints

Most state constrained stationary problems can be treated directly with interior point techniques. We gained a thorough analytic understanding of the effect of barrier regularizations and their corresponding homotopies. In particular, the homotopy path enjoys simple and strong convergence and continuity properties. Discretization schemes for central path solutions were developed and analyzed. It turned out that the use of appropriate rational barrier functions, as opposed to the classical logarithmic barrier function, leads to a strictly feasible path. This in turn, is the basis for a convergence theory of a Newton path-following method, a result which is up to now unique in the literature. For a finite dimensional control space, we could determine an optimal rational order, depending on the spatial dimension of the problem, which leads to linear convergence in generic situations. The arrier approach was extended to a problem with maximum-norm objective. This setting is of special interest in the formulation of optimization criteria in local hyperthermia. Further, we extended our barrier techniques to gradient bounds on the state.

Interior point methods for time-dependent problems with state constraints

Based on the expertise obtained for elliptic control problems, particular emphasis was laid on the construction and analysis of solution algorithms for parabolic problems. These will gain particular importance in the context of transient and periodic heating strategies. In most relevant cases additional Lavrentiev regularization is necessary. A conceptual primal interior point method was developed for time-dependent problems with pointwise state constraints. After regularization, convergence was shown for a logarithmic barrier function. Here, adaptive time discretization techniques were decisive to deal with irregular Lagrange multipliers. The analysis is similar to the one for the elliptic case, but in most time dependent problems continuity of the states is not guaranteed and thus a classical theoretical foundation is not available. This obstacle is completely overcome by the Lavrentiev type regularization, which permits to derive KKT-conditions for any dimension of the domain.

Function space oriented optimization of treatment parameters

Up to now, superposition of temperature fields corresponding to the antenna parameters had been exploited for fast and easy optimization of the antenna parameters. This scheme depends inherently on the linearity of the BHTE and fails if more accurate nonlinear perfusion models are used. Moreover, precomputation of the temperature fields must be done without knowledge about the superposed temperature field, such that adaptive mesh refinement oriented towards the final solution is impossible. In addition, the temperature constraints, which have been treated by an outer penalty approach, should be treated as inequality constraints.

Heat distribution after optimization.

Algorithmic development

Based on the analysis of the barrier homotopy path an adaptive path-following algorithm was constructed and analyzed. The algorithm exploits the special problem structure by a pointwise damping step, which increases the efficiency and robustness of the algorithm significantly, compared to the standard variant.

ALG
Choice of step-sizes of interior point algorithms with and without pointwise damping

An implementation has been applied successfully to hyperthermia treatment planning problems. In particular it has been shown to retain its good numerical performance also for application problems with complex geometries and irregular data. Currently, we are extending this solver for clinical use. A powerful approach to exploit the function space structure of our problem are inexact Newton path-following methods in function space. This approach integrates non-linear algorithms and adaptive refinement by a-posteriori error control of the Newton steps. The steps are performed inexactly in function space. A crucial issue is the choice of an appropriate scaled norm, which governs the convergence behavior of Newton's method. We were able to construct an a-posteriori error estimator fits exactly into this framework. The resulting algorithms shows excellent numerical behavior, since most of the iterations can be performed on coarse computational grids. This happens, although the radius of convergence of the Newton corrector shrinks to zero in the limit of the homotopy. A promising variant of goal-oriented error estimation and mesh adaptivity for optimization problems based on the actual cost reduction has been developed. First results show a significant improvement over existing approaches. One possibility to address optimal control problems is to apply standard solvers to discretized problems. The NLP code IpOpt has been used to solve hyperthermia treatment problems on fixed discretizations. The performance has not been satisfying compared to the function space oriented code developed in A1.

Left: Optimal control for a model problem. Right: Adaptively chosen grid.

The research on interior point methods was complemented by a new numerical approach for solving optimality systems in state-constrained PDE control. Here, a Lavrentiev type regularization and control reduction are appropriate tools to convert the optimality system into a Newton-differentiable coupled system of nonlinear PDEs that can be solved by available codes for PDEs. This method was successfully applied to parabolic problems up to space dimension 2 and nonlinear elliptic problems of dimension 3 related to regional hyperthermia.

[back to the top]

Publications

U. Prüfert, A. Schiela

The Minimization of a maximum-norm functional subject to an elliptic PDE and state constraints
Z. angew. Math. Mech., Vol. 89 (7) , 536-551 (2009)
Abstract Preprint
A. Schiela

State constrained optimal control problems with states of low regularity
SIAM J. on Control and Optimization, Vol. 48 (4) , 2407-2432 (2009)
Abstract Preprint
A. Schiela, A. Günther

Interior Point Methods in Function Space for State Constraints - Inexact Newton and Adaptivity
(09-01) ZIB Report (2009)
Abstract Preprint
A. Schiela

Barrier Methods for Optimal Control Problems with State Constraints
SIAM J. on Optimization -to appear (2009)
Abstract Preprint
A. Schiela

An Extended Mathematical Framework for Barrier Methods in Function Space
In: Domain Decomposition Methods in Science and Engineering XVIII, Lecture Notes in Computational Science and Engineering, Vol. 70 (2009)
Abstract Preprint
M. Weiser

Optimization and Identification in Regional Hyperthermia
Int. J. Appl. Electromagn. and Mech, Vol. 29, 1–11 (2009)
Abstract Preprint
M. Hinze, A. Schiela

Discretization of Interior Point Methods for State Constrained Elliptic Optimal Control Problems: Optimal Error Estimates and Parameter Adjustment.
Comp. Opt. and Appl. -to appear (2009)
Abstract Preprint
A. Schiela, M. Weiser

Superlinear convergence of the Control Reduced Interior Point Method for PDE Constrained Optimization
Computational Optimization and Applications, Vol. 39 (3) , 369-393 (2008)
Abstract Preprint
U. Prüfert, F. Tröltzsch, M. Weiser

The convergence of an interior point method for an elliptic control problem with mixed control-state constraints
Comput. Optim. Appl., Vol. 39 (2) , 183–218 (2008)
Abstract Preprint
O. Schenk, A. Wächter, M. Weiser

Inertia Revealing Preconditioning For Large-Scale Nonconvex Constrained Optimization
SIAM J. Sci. Comp., Vol. 31 (2) , 939–960 (2008)
Abstract Preprint
A. Schiela, W. Wollner

Barrier Methods for Optimal Control Problems with Convex Nonlinear Gradient Constraint
(08-47) ZIB Report (2008)
Abstract Preprint
A. Schiela

An Interior Point Method in Function Space for the Efficient Solution of State Constrained Optimal Control Problems
(07-44) ZIB Report (2008)
Abstract Preprint
I. Schräer

Modellreduktionï¿½inï¿½derï¿½Hyperthermieï¿½mittelsï¿½POD
FU Berlin, Fachbereich Mathematik Master's thesis (2008)
R. Griesse, M. Weiser

On the Interplay Between Interior Point Approximation and Parametric Sensitivities in Optimal Control
Journal of Mathematical Analysis and Applications, 771-793 (2008)
Abstract Preprint
S. Volkwein, F. Tröltzsch

POD a-posteriori error estimates for linear-quadratic optimal control problems
(SFB-Report 2007-008) SFB Research Center "Mathematical Optimization and Applications in Biomedical Sciences" (2007)
U. Prüfert, F. Tröltzsch

An interior point method for a parabolic optimal control problem with regularized pointwise state constraints
ZAMM, Vol. 87 (8–9) , 564–589, Berlin (2007)
URL
C. Meyer, U. Prüfert, F. Tröltzsch

On two numerical methods for state-constrained elliptic control problems
Optimization Methods and Software, Vol. 22 (6) , 871–899 Preprint (2007)
URL
I. Neitzel, U. Prüfert, T. Slawig

Strategies for time-dependent PDE control using an integrated modeling and simulation environment. Part one: problems without inequality constraints
(408) Matheon, Berlin (2007)
URL
M. Weiser, P. Deuflhard

Inexact central path following algorithms for optimal control problems
SIAM J. Control Opt., Vol. 46 (3) , 792–815 (2007)
Abstract Preprint
M. Weiser, P. Deuflhard, B. Erdmann

Affine conjugate adaptive Newton methods for nonlinear elastomechanics
Opt. Meth. Softw., Vol. 22 (3) , 413-431 (2007)
I. Neitzel, U. Prüfert, T. Slawig

Strategies for time-dependent PDE control using an integrated modeling and simulation environment. Part two: problems with inequality constraints
(24-02) SPP 1253 (2007)
M. Weihrauch, P. Wust, M. Weiser, J. Nadobny, S. Eisenhardt, V. Budach, J. Gellermann

Adaptation of antenna profiles for control of MR guided hyperthermia (HT) in a hybrid MR-HT system
Medical Physics, Vol. 34 (12) , 4717-4725 (2007)
Revised Version
M. Weiser, T. Gänzler, A. Schiela

A control reduced primal interior point method for a class of control constrained optimal control problems
Comput. Optim. Appl., Vol. 41 (1) , 127-145 (2007)
Abstract Preprint
A. Schiela

Optimality Conditions for Convex State Constrained Optimal Control Problems with Discontinuous States
(07-35) ZIB Report (2007)
Abstract Preprint
E. Casas, J. C. de Los Reyes, F. Tröltzsch

Sufficient second-order optimality conditions for semilinear control problems with pointwise state constraints
(2007)
Preprint
S. Gerber

Perfusionsmodellierung in menschlichen Tumoren
FU Berlin, Fachbereich Mathematik Master's thesis (2007)
A. Massing

1D-Reduktion thermal signifikanter Aderstränge in der Hyperthermie-Modellierung
FU Berlin, Fachbereich Mathematik Master's thesis (2007)
S. Zachow, H.-C. Hege, P. Deuflhard

Computer Assisted Planning in Cranio-Maxillofacial Surgery
Journal of Computing and Information Technology - Special Issue on Computer-Based Craniofacial Modelling and Reconstruction (2006 (to appear))
P. Deuflhard, M. Weiser, S. Zachow

Mathematics in Facial Surgery
AMS Notices, Vol. 53 (9) , 1012–1016 (2006)
T. Gänzler, S. Volkwein, M. Weiser

SQP methods for parameter identification problems arising in hyperthermia
Optim. Methods Softw., Vol. 21 (6) , 869–887 (2006)
URL
A. Schiela

A continuity result for Nemyckii Operators and some applications in PDE constrained optimal control
(06-41) ZIB ZIB Report (2006)
Abstract Preprint
A. Schiela

Convergence of the Control Reduced Interior Point Method for PDE Constrained Optimal Control with State Constraints
(06-16) ZIB (2006)
Abstract Preprint
P. Deuflhard

Newton Methods for Nonlinear Problems. Affine Invariance and Adaptive Algorithms
Springer, Springer Series in Computational Mathematics -2. überarb. u. erweit. Aufl. (2006)
Outline
A. Schiela

The Control Reduced Interior Point Method. A Function Space Oriented Algorithmic Approach
FU Berlin, Fachbereich Mathematik Doctoral thesis (2006)
J. Gellermann, M. Weihrauch, C. H. Cho, W. Wlodarczyk, H. Fähling, R. Felix, V. Budach, M. Weiser, J. Nadobny, P. Wust

Comparison of MR-thermography and planning calculations in phantoms
Med. Phys., Vol. 33, 3912–3920 (2006)
URL
M. Weiser

Interior point methods in function space
SIAM J. Control Optimization, Vol. 44 (5) , 1766-1786 (2005)
Abstract Preprint Revised Version
M. Weiser, A. Schiela, P. Deuflhard

Asymptotic Mesh Independence of Newton's Method Revisited
SIAM J. Num. Anal., Vol. 42 (5) , 1830-1845 (2005)
Abstract Preprint Revised Version
F. Tröltzsch

Optimale Steuerung partieller Differentialgleichungen. Theorie, Verfahren und Anwendungen
Vieweg (2005)
S. Zachow, M. Weiser, H.-C. Hege, P. Deuflhard

Soft Tissue Prediction in Computer Assisted Maxillofacial Surgery Planning: A Quantitative Evaluation of Histomechanical Modeling using Pre- and Postoperative CT-Data
In: Biomechanics Applied to Computer Assisted Surgery, Payan, Y. (ed) Research Signpost, 277-298 (2005)
P. Deuflhard, R. Hochmuth

Multiscale analysis of thermoregulation in the human microvascular system
Wiley VerlagMath. Methods Appl. Sci., Vol. 27 (8) , 971-989 (2004)
R. Hochmuth, P. Deuflhard

Multiscale Analysis for the Bio-Heat-Transfer Equation - The Nonisolated Case
Mathematical Models and Methods in Applied Sciences (M^3AS), Vol. 14 (11) , 1621-1634 (2004)
M. Weiser, A. Schiela

Function space interior point methods for PDE constrained optimization
PAMM, Vol. 4 (1) , 43-46 (2004)
Abstract Preprint
S. Zachow, H.-C. Hege, P. Deuflhard

Computergestützte Operationsplanung in der Gesichtschirurgie
In: Proc. VDE Kongress 2004 - Innovation für Menschen, Band 2, Fachtagungsberichte DGBMT - GMM - GMA, 53-58 (2004)
J. Nadobny, R. Pontalti, D. Sullivan, W. Wlodarczyk, A. Vaccari, P. Deuflhard, P. Wust

A Thin-Rod Approximation for the Improved Modeling of Bare and Insulated Cylindrical Antennas Using the FDTD Method
IEEE Transactions on Antennas and Propagation, Vol. 51 (8) , 1780-1796 (2003)
J. Nadobny, D. Sullivan, W. Wlodarczyk, P. Deuflhard, P. Wust

A 3-D Tensor FDTD-Formulation for Treatment of Sloped Interfaces in Electrically Inhomogeneous Media
IEEE Transactions on Antennas and Propagation, Vol. 51 (8) , 1760-1770 (2003)
G. Sreenivasa, J. Gellermann, J. Nadobny, P. Schlag, P. Deuflhard, R. Felix, P. Wust

Clinical Use of the Hyperthermia Treatment Planning System Hyperplan to Predict Effectiveness and Toxicity
Int. J. Oncology Biol. Phys., Vol. 55 (2) , 407-419 (2003)
R. Hochmuth, P. Deuflhard

On the thermoregulation in the human microvascular system
PAMM, Vol. 3, 378-379 (2003)
C. Kober, R. Sader, H. F. Zeilhofer, P. Deuflhard

An Individual Anisotropic and Inhomogeneous Material Description of the Human Mandible
In: Mathematical Modelling and Computing in Biology and Medicine, V. Capasso (ed) ESTMB, 103-109 (2003)
J. Nadobny, W. Wlodarczyk, P. Deuflhard, G. Mönich, P. Wust

Numerical and Experimental Investigations of Different Antenna Types for Deep Body Hyperthermia
In: Proc. 2001 URSI International Symposium on electromagnetic Theory, M.A. Stuchly and D.G. Shannon (ed) , 473-475 (2001)