Plan:
To bother about the best method of accomplishing an accidental
result.
Ambrose Bierce, The Enlarged
Devil’s Dictionary
Single and multiobjective anatomy-based dose optimization for HDR Brachytherapy
WinOpt-HDR is a toolkit for multiobjective anatomy based dose optimization in HDR-brachytherapy.
This demo version is limited to the post implant optimization, the inverse planning module is not supported. Run setup.exe to install WinOpt-HDR. Some test data are included in this demo. Before downloading this demo see the PDF file (4 MB) for more information about this demo toolkit.
If you are interested send an e-mail to mlahanas@gmx.net to get the password required for the installation of this demo toolkit. The author is interested of who and why is interested in this program and also hopes to receive some comments about this demo.
The algorithms have been further improved resulting in a commercial treatment planning system.
A revolutionary new concept for HDR prostate treatment
Multiobjective evolutionary algorithms are now used every day in clinical practice in radiotherapy using SWIFT, a new system for HDR prostate treatment, developed by Pi-Medical and MedCom, now available by Nucletron.
The Nucletron SWIFT™ system provides the user with a dedicated 3D planning system for HDR Prostate Treatment. Transversal ultrasound images are captured, using a standard stepper unit and ultrasound system, and transformed into 3D information. This is used to determine the optimal implant geometry and to guide the user during the insertion of the needles. The optimal treatment plan is created while the patient is still in the operating room!
Pre-Plan / Live Plan / Optimization
Nucletron SWIFT™ is a unique and exciting brachytherapy treatment planning system ideal for Prostate treatment using the microSelectron© HDR remote afterloading system. The new system combines Nucletron's strengths as a world leader in brachytherapy with the convenience of Ultrasound Imaging to improve the speed and effectiveness of the prostate cancer treatment. New algorithms, optimized for speed and accuracy, and stepper enhancements enable the user to acquire the 3D ultrasound information and calculate the necessary optimization and evaluation information within seconds. Nucletron SWIFT™ is the fastest route to the first fraction!
If you want to receive more information or if you like a Nucletron representative contact you for a SWIFT demonstration, click here
Download from the link below:
Contributors to this demo optimization program:
Windows
98, NT, ME, 2000
Microsoft Corporation (Thanks for these beautiful operating systems!)
Triangulation
H. Fuchs. Z. M. Kedem, S. P. Uselton, M. Manocha, A. Narkhede , D. Manocha, S. Giannouli, K. Karouzakis
OPENGL
Silicon Graphics, Microsoft
2D and 3D
Isodose distributions
W.E. Lorensen, H. E. Cline
Look-Up
tables for speed-up
M. Lahanas, K. Karouzakis
Dosimetric-Look-Up
tables
P. Karaiskos, Prof. D. Baltas, M. Lahanas, M. Papagiannopoulou, K. Karouzakis, F. Lliso, J. Perez-Calatayud, V. Carmona
TIFF and
BMP
Microsoft Corporation, U. Ackermann, M. Lahanas
Multiobjective
Optimization
K. Deb, M. Lahanas, K. Karouzakis, E. Zitzler, P. Sevinc, N. Milickovic, J. Knowles, D. Corne, A. Jaszkiewicz
Deterministic
Optimization
W. H. Press, S. A. Teukolsky, W. T. Vetterling, B. P. Flannery
Simulated
Annealing
Taygeta Software
Histogram
Classes
K. Karouzakis
Inverse
Planning
K. Karouzakis, M. Lahanas, Prof. D. Baltas, D. Eberly
Geometric
Optimization
(Optimized bounding boxes, virtual polygons etc., sampling point generation) M. Lahanas, D. Eberly, K. Karouzakis, S. Giannouli, T. Kemmerer
FFT Dose
Calculation
M. Frigo, S. G. Johnson, T. Kemmerer, M. Lahanas
Catheter
Reconstruction
N. Milickovic, Prof. D. Baltas
Dose
Calculation
Prof. D. Baltas, M. Papagiannopoulou, M. Lahanas
Collision
Detection
S. Gottschalk
Objective
functions
Prof. D. Baltas, J. Pouliot, M. Lahanas, M. Papagiannopoulou
Decision
Making, Filter
M. Lahanas, Prof. D. Baltas, K. Karouzakis
Datasets
Prof. D. Baltas, N. Milickovic, S. Giannouli and Medical Physics and Engineering Department, Klinikum Offenbach, Germany
Dose
Calculation
Prof. D. Baltas, M. Papagiannopoulou, M. Lahanas
Plot
Classes
R. Hall
Optimization
Tests
M. Papagiannopoulou, Prof. D. Baltas
Financial
Support
Prof. D. Baltas, MITTUG- IST-1999-10618 Project
Purpose
of this demo version
Optimization in HDR
brachytherapy
Today the majority of treatment planning systems in brachytherapy such as the Nucletron PLATO system use phenomenological optimization methods such as geometrical optimization. Additionally most of the algorithms used have the so called artificial problem of negative times which in principle does not exist and artificial methods such as setting the negative dwell times equal to zero and applying a dose renormalization are therefore used. 20%-50% of the dwell times that are always negative as a result of the optimization are arbitrarily set equal to 0. Dose normalization is then applied to rescale the dose at a specified number of dose points.
Another method used is to introduce constraints on the objective function such as gradient constraints between dwell times of adjacent dwells, which reduces but does not completely avoid negative solutions. Still up to 20% of the dwell times can be negative. There is no reason why a gradient-based restriction should be applied to the objective function only to reduce negative times which can be avoided by a more natural and extremely simple method without imposing any restriction on the optimization result. Other approaches use a constraint of positive dwell times. They do not require some rather arbitrary modification of the objective function. This in principle increases the number of optimization parameters by a factor of two. If the constraints are satisfied only partial and negative dwell times are still present in the final solution again a correction mechanism must be applied.
An additional constraint applied on the optimization result is the use of a dose point normalization based on the dose of some points for example on the minimum peripheral dose, or the average dose on the PTV surface. This is because the definition of what conformal brachytherapy means is not mathematical well defined. If the most important objective is to have a dose at least above some value in the entire PTV then a normalization based on the minimum peripheral dose can be used. By using this normalization method it is true that the dose in the PTV will be everywhere high enough but how is then the dose distribution in the surrounding tissue and in organs at risk? This method based on the normalization on a single point usually increases the dose in the surrounding tissue by in principle expanding the reference isodose so that it passes through this dose point at which the minimum dose was measured and if the point is on the PTV surface. If the dose is normalized using the PTV dose average value then we cannot have always high enough dose values anywhere in the PTV. As long as there are no sources outside the PTV it will more or less help us to avoid high dose values in the surrounding tissue. Also we have not full control of the dose in the organs at risk. This method assumes that the reference isodose can actually have the shape of the PTV. As the shape of the PTV is usually complicated and isodose surfaces are less complex it requires that the number of sources is large and adequately distributed in the PTV.
A method that is less restricted is the normalization free method. In this method there is in principle no point defined at which the dose is normalized. The optimization is free so that finally we have the best possible results. Currently there are some requirements to document the dose in a definite way and the established methods introduce constraints that reduce the quality of the dose distribution which can be obtained.
It is still not clear that the problem is not a single objective problem but a multiobjective problem with contradictory and competing objectives that requires the determination of at least a portion of the Pareto front. Today when there are no exact established quality criteria the planner cannot guide the system exactly to a single solution on the Pareto front.
The problem is transformed into a single objective problem using arbitrary importance factors for each objective and combining the objectives into a single objective function to be minimized. What does it mean to use an importance factor 0.9 for objective 1 and 0.1 for objective 2? That the objective 1 is more or less 9 times more important than objective 2? If there is a strong tradeoff then a combination of 0.89 and 0.11 can give quite different results! Some algorithms require importance factors of 20000 for one objective and 10 for another which shows that also the optimal importance factors depend on the objective function used. In IMRT we have found that the surrounding tissue and PTV coverage objective produce not the best results with a 0.1 and 0.9 importance factor respectively as assumed but sometimes the best result was obtained with importance factors 0.7 and 0.3 or 0.3 and 0.7 respectively. This is because some objectives are cooperating to some extent and to some extent they are competing, depending on the importance factors of the other objectives and in a complicated way which depends on the topology of the PTV and the OARs, the beam orientation etc. It is not very much different in brachytherapy.
Not only is the optimization method important but also the distribution of sampling points used in the optimization procedure. One problem is that dose variances are calculated from sampling points limited on the contours which even if their number is large it is not true that the variances are approximated with sufficient accuracy. Another problem is the use of points inside catheters.
Optimization is the most important part of a treatment planning system and to a smaller extent for example its volume rendering capabilities. It seems therefore strange that these constraints used in the optimization, models based on crude approximations are still used in the majority of planning systems.
It is true that speed is an important factor. Nevertheless also methods such a geometric optimization although may be fast require from the planners often to manually intervene in order to modify the dwell times if the results are not satisfactorily. Using only a very small set of importance factors, 1-2 usually, the planner is left without information of what actually is possible. The true multiobjective optimization is free of importance factors. Even if a set of importance factors which somehow gives reasonable results can be found for some fixed topology it requires time to find these and a significant time has to be spent. A multiobjective algorithm does not require such training.
Sometimes although it seems strange in principle it is easier to calculate all possible solutions at once than only one particular. This is true for true multiobjective optimization algorithms. In a single objective optimization there is only a single decision, accept the solution or not. If not a new solution must be calculated or a manual manipulation of dwell times or dose rescaling remains which takes a lot of time, usually much more than the automatic optimization.
These are definitive not methods of the 21st century. Today the computing power of modern PCs allows the determination of hundreds of possible solutions. The planner now has information that was to him/her not available before or only to a very limited extent. The optimization problem is not an optimization problem anymore that has to consider arbitrary importance factors, arbitrary correction methods and objectives based on nomograms or other hand waving arguments, but are now transformed to a decision making problem. The planner is confronted with all possible solutions and information. He/She knows now all the answers to the questions: What is possible if I modify the importance factors? What coverage can I get? How this affects the dose value in the organs at risk? Even if some new systems give an approximate answer to some of these questions these methods use a sometimes-complicated way to obtain part of this information.
A true multiobjective method requires a set of objectives functions that are more intuitive for the planner than variances of dose distributions. The most natural way from a dosimetric based approach seems to be the use of dose-volume histogram based objectives.
Methods have been proposed which try to optimize the importance factors using an optimization method. The objective function for a set of importance factors is calculated and then a different set of functions combined with another set of new importance factors is used to establish a quality criterion of the optimization results. Here the problem has been bypassed by the introduction of a new set of artificial importance factors, but this method is probably better than using a fixed set of arbitrary importance factors for the objective functions.
It is clear that methods such as the geometric optimization method have been established through the years. The hope is that the new optimization methods slowly will be introduced in new planning systems and the planners will get familiar with the new methods. It requires some training but the result will be less arbitrary without unnecessary correction methods, less manual intervention. It will also produce better optimization results. Finally it will produce probably new quality criteria that at the moment are not very well established. It is true that we cannot expect miracles from any optimization method since the number of sources, and their distribution finally introduces physical restrictions of what is possible. Aim of a true multiobjective anatomy dose optimization is to show these limitations independent on arbitrary corrections and importance factors. It is also necessary not to use the word inverse planning for only post plan optimization, since inverse planning means actually to find not only the dwell times of the optimal solution but also the optimal distribution of sources in space in an economic way, i.e. finding the smallest possible number of source dwell positions without significantly reducing also the optimization quality in comparison to a solution a very large number (within realistic limits) of sources.
It should become clear soon or later that only multiobjective optimization should be used for the inverse planning problem.
The new optimization methods do not require normalization on some arbitrary set of dose points. They do not require arbitrary correction mechanisms. They do not require arbitrary importance factors. All these constraints in the past have due to ignorance or due to limiting computational possibilities introduced restrictions on the optimization result. The dominance of the market by a single provider for treatment planning systems in brachytherapy has helped to establish the old methods and it is now difficult to replace these by the new methods.
It is true that the new optimization methods require a slightly more complex cooperation with planners and still the decision making process of these multiobjective methods can be improved with some additional tools. A dose optimization procedure which does not require any human decision making process is difficult if not impossible to be realized.
References for some of the methods used in WinOpt-HDR
M. Lahanas, D. Baltas and E. Schreibmann. Application of Multiobjective Evolutionary Algorithms for radiotherapy, submitted to Applied Soft Computing 2003
M. Lahanas, K. Karouzakis, S. Giannouli, R. F. Mould and D. Baltas. Inverse Planning in Brachytherapy: Radium to High Dose Rate 192 Iridium Afterloading, to be published as invited review in Nowotwory Journal of Oncology 2003
14) M. Lahanas, D. Baltas and S. Giannouli, Global Convergence Analysis of Fast Multiobjective Gradient based Dose Optimization Algorithms for High Dose Rate Brachytherapy, Phys. Med. Biol. 48 599-617, 2003
M. Lahanas, D. Baltas and N. Zamboglou. A Hybrid Evolutionary Multiobjective Algorithm for Anatomy-Based Dose Optimization Algorithm in HDR Brachytherapy, Phys. Med. Biol. 48 399-415, 2003
K. Karouzakis, M. Lahanas, N. Milickovic, S. Giannouli, D. Baltas and N. Zamboglou. Brachytherapy dose–volume histogram computations using optimized stratified sampling methods, Med. Phys. 29 424-432, 2002
T. Kemmerer, M. Lahanas, D. Baltas and N. Zamboglou, DVH computation comparisons using conventional methods and optimized FFT algorithms for brachytherapy, Med. Phys. 27, 2343-2356, 2000.
M. Lahanas, D. Baltas, N. Milickovic, S. Giannouli, and N. Zamboglou, Generation of uniformly distributed dose points for anatomy-based- three-dimensional dose optimization in brachytherapy, Med. Phys. 27, 1034-1046 2000
M. Lahanas, D. Baltas and N. Zamboglou, Anatomy-based three-dimensional dose optimization in brachytherapy using multiobjective genetic algorithms, Med. Phys. 26 1904-1918, 1999
M. Lahanas, T. Kemmerer, N. Milickovic, D. Baltas, N. Zamboglou, Optimized bounding boxes for three-dimensional treatment planning in brachytherapy, Med. Phys. 27, 2333-2342, 2000.
M. Lahanas, N. Milickovic, D. Baltas and N. Zamboglou, Application of Multiobjective Evolutionary Algorithms for Dose Optimization Problems in Brachytherapy, in Proceedings of the first international conference, EMO 2001, Zurich, Switzerland, edited by E. Zitzler, K. Deb, L. Thiele, C. A. Coello Coello, D. Corne, Lecture Notes in Computer Science Vol. 1993, Springer, 574-587, 2001
N. Milickovic, M. Lahanas, D. Baltas and N. Zamboglou, Comparison of Evolutionary and Deterministic Multiobjective Algorithms for Dose Optimization in Brachytherapy, in Proceedings of the first international conference, EMO 2001, Zurich, Switzerland, edited by E.Zitzler, K. Deb, L. Thiele, C. A. Coello Coello, D. Corne, Lecture Notes in Computer Science Vol. 1993, Springer 167-180, 2001.
M. Lahanas, N. Milickovic, M. Papagiannopoulou, K. Karouzakis, D. Baltas and N. Zamboglou, Application of a Hybrid NSGA-II Multiobjective Algorithm for Anatomy based Dose Optimization in Brachytherapy, "EUROGEN 2001 - Evolutionary Methods for Design, Optimisation and Control with Applications to Industrial Problems" Athens, Greece 19-21 September 2001.
N. Milickovic, M. Lahanas, M. Papagiannopoulou, K. Karouzakis, D. Baltas and N. Zamboglou, Application of Multiobjective Genetic Algorithms in Anatomy based Dose Optimization in Brachytherapy and its Comparison with Deterministic Algorithms, "EUROGEN 2001 - Evolutionary Methods for Design, Optimisation and Control with Applications to Industrial Problems" Athens, Greece 19-21 September 2001.
N. Milickovic, M. Lahanas, M. Papagiannopoulou, K. Karouzakis, D. Baltas and N. Zamboglou, Application of Multiobjective Genetic Algorithms in Anatomy Based Dose Optimization in Brachytherapy and its Comparison with Deterministic Algorithms, Proceedings of the 23rd IEEE EMBS Int. Conference, 23rd Annual International Conference of the IEEE - Engineering in Medicine and Biology Society, Istanbul, Turkey, 25-28 October, 2001.
In preparation
M. Lahanas, Application of Multiobjective Evolutionary Optimization Algorithms in Medicine, in Application of Multiobjective Evolutionary Optimization Algorithms, edited by C. Coello Coello et al, World Scientific 2004
Update: 1-11-2003