Dose-volume constrained optimization in intensity-modulated radiation therapy treatment planning
- Yoshihiro Tanaka^{1},
- Ken’ichi Fujimoto^{2} and
- Tetsuya Yoshinaga^{2}Email author
https://doi.org/10.1186/s13660-015-0643-2
© Tanaka et al.; licensee Springer. 2015
Received: 29 October 2014
Accepted: 26 March 2015
Published: 8 April 2015
Abstract
We present a novel optimization method to handle dose-volume constraints (DVCs) directly in intensity-modulated radiation therapy (IMRT) treatment planning based on the idea of continuous dynamical methods. Most of the conventional methods are constructed for solving inconsistent inverse problems with, e.g., dose-volume based objective functions, and one expects to obtain a feasible solution that minimally violates the DVCs. We introduce the concept of ‘acceptable’, meaning that there exists a nonempty set of radiation beam weights satisfying the given DVCs, and we resolve the issue that the objective and evaluation are different in the conventional planning approach. We apply the initial-value problem of the proposed dynamical system to an acceptable and inconsistent inverse problem and prove that the convergence to an equilibrium in the acceptable set of solutions is theoretically guaranteed by using the Lyapunov theorem. Indeed, we confirmed that we can obtain acceptable beam weights through numerical experiments using phantom data simulating a clinical setup for an acceptable and inconsistent IMRT planning system.
Keywords
intensity-modulated radiation therapy treatment planning dose-volume constraints dose-volume constrained optimization inverse problem stability of solution1 Introduction
Intensity-modulated radiation therapy (IMRT) [1, 2] is an effective radiotherapy technique that attacks cancer without inflicting serious damage to critical normal tissues by delivering radiation beams from many angles. For designing IMRT inverse plans, an optimization strategy is typically used to minimize an objective or cost function of radiation beam weights [3]. We widely utilize the objective function to achieve as high as possible proportions of the total volumes receiving doses greater than and less than the prescribed doses for planning target volumes (PTVs) and organs at risk (OARs), respectively. However, in practice, it is often impossible to fill all the volumes with permissible doses. Namely, the inverse problem generally becomes inconsistent, so the optimization problem has no optimum solution. Dose-volume constraints (DVCs) expressed as a percentage of the prescription dose are a natural way to specify the objective and have been the standard way of evaluating the treatment in practice. Most of the conventional methods are constructed for solving inconsistent inverse problems with, e.g., dose-based, dose-volume based, and biology-based objective functions [4–8], and one expects to obtain a feasible solution that minimally violates the DVCs. Therefore, planning for IMRT requires a trial-and-error process and human intervention [9] due to the difference between the optimization objective and the evaluation of planning results.
In this paper, we introduce the concept of ‘acceptable’ meaning that there exists a nonempty set of radiation beam weights satisfying the given DVCs on PTVs and OARs in the IMRT planning system and resolve the issue that the objective and evaluation are different in the conventional planning approach. That is, we present a novel optimization method to handle DVCs directly in IMRT treatment planning based on the idea of continuous dynamical methods [10–20]. We apply the initial-value problem of a dynamical system, which is described by nonlinear differential equations for an acceptable and inconsistent inverse problem, and prove that the convergence to an equilibrium [21] in the acceptable set of solutions is theoretically guaranteed by using the Lyapunov theorem [22]. Then we can obtain a set of acceptable beam weights over time with decreasing Kullback-Leibler divergence [23, 24] measure. We also show the fact that the intensities of all radiation beams are not negative and are less than a specific upper limit, which is achieved in accordance with a box-constrained optimization procedure. The proposed dynamical system extends our previously presented nonlinear continuous-time system [20] for solving a split feasibility problem [25–27]. A simulated dynamics approach [28] also uses a differential equation, but it is difficult to ascertain whether its solution has converged to the global optimum.
We examined numerical experiments using phantom data simulating a clinical setup for an acceptable and inconsistent IMRT planning system. The system proposed in this paper is compared with another dynamical system constructed for optimizing a consistent inverse problem. We found that the proposed system can provide an acceptable solution, but the other cannot.
2 IMRT planning
Let \(D_{1}^{\mathrm{L}}\) and \(D_{2}^{\mathrm{U}}\) represent the lower and upper bounds of doses for PTV and OAR, respectively (\(D_{1}^{\mathrm{L}}> 0\) and \(D_{2}^{\mathrm{U}}\ge0\)). Additionally, we define an upper dose bound for PTV as \(D_{1}^{\mathrm{U}}> D_{1}^{\mathrm{L}}\) to avoid excessively high dose values inside the PTV.
Definition 1
Definition 2
We define the IMRT planning system to be acceptable if there exists a common beam set such that dose distributions in PTVs and OARs are partly acceptable for all DVCs.
Definition 3
Obviously, if the IMRT planning system is consistent, it is acceptable. We are interested in the situation where the system is inconsistent and acceptable. In this paper, the problem of dose-volume constrained optimization in IMRT planning is defined to obtain the unknown variable \(x \in{\mathcal{A}}\) if the system is acceptable.
3 Dynamical system for dose-volume constrained optimization
We give theoretical results for the behavior of the solution to the dynamical system in Eq. (8). First, we show that all solutions stay inside the hypercube.
Proposition 1
If we choose the initial value \(x^{0} \in\Omega\) in the dynamical system in Eq. (8), then the solution \(\varphi(t, x^{0})\) stays in Ω for all \(t > 0\).
Proof
Since the system can be written as \(dx_{j}/dt = - x_{j} (1 - \gamma^{-1} x_{j}) (K^{\top})_{j} Q(\operatorname {Log}(K x) - \operatorname{Log}(P(K x)))\), we see that for any j the solution satisfies \(d\varphi_{j}/dt \equiv0\) on the subspace where \(x_{j} = 0\) or \(x_{j} = \gamma\). Therefore, the subspace is invariant, and trajectories cannot pass through every invariant subspace in accordance with the uniqueness of solutions for the initial-value problem. This leads to any solution \(\varphi(t, x^{0})\) of the system in Eq. (8) with initial value \(x^{0} \in\Omega\) being in Ω for all \(t > 0\). □
Next, we prove the stability of an equilibrium in the set \({\mathcal {A}}\), which corresponds to the desired radiation beam weights. Namely, the existence of a Lyapunov function for the system in Eq. (8) guarantees the stability of the equilibrium set [21].
Theorem 1
If the IMRT planning system is acceptable, then the equilibrium set of Eq. (8) is stable.
Proof
The theoretical results show that an element in the acceptable set can be obtained by applying the initial-value problem of the hybrid dynamical system with piecewise continuous vector fields in Eq. (8). To be acceptable, the vector of beams or the solution \(\varphi(t, x^{0})\) must behave appropriately for the volume percentage in the dose-volume constraints, roughly speaking, in the following manner. Let us assume, e.g., the dose \(D_{1}= K_{1} \varphi(t, x^{0})\) for PTV does not satisfy the upper bound constraint \((D_{1}^{\mathrm{U}},\zeta_{1}^{\mathrm{U}})\), i.e., it is not partly acceptable. The solution \(\varphi(t, x^{0})\) causes the dose to change its distributions along the gradient of the Kullback-Leibler divergence measure so that \(D_{1}\) can satisfy the constraint. The same manner is applied to the lower and upper bounds of doses for PTV and OAR, respectively. When all dose distributions in PTVs and OARs become partly acceptable, the vector field is entirely zero after the solution has converged to an equilibrium.
4 Materials for numerical experiments
Dose-volume constraint or equivalent parameters for partly acceptable dose distribution in our example
Assigned region (color in Figure 1 ) | Organ | Dose-volume constraint | Set of parameters for being partly acceptable |
---|---|---|---|
PTV (blue) | Prostate | \(V_{82.1} \ge87\%\) | \((D_{1}^{\mathrm{L}},\zeta_{1}^{\mathrm{L}}) = (82.1, 0.87)\) |
\(V_{95.9} < 2\%\) | \((D_{1}^{\mathrm{U}},\zeta_{1}^{\mathrm{U}}) = (95.9, 0.98)\) | ||
OAR1 (green) | Right femoral head | \(V_{30} < 10\%\) | \((D_{2}^{\mathrm{U}},\zeta_{2}^{\mathrm{U}}) = (30, 0.9)\) |
OAR2 (red) | Bladder | \(V_{20} < 60\%\) | \((D_{2}^{\mathrm{U}},\zeta_{2}^{\mathrm {U}}) = (20, 0.4)\) |
\(V_{45} < 30\%\) | \((D_{2}^{\mathrm{U}},\zeta_{2}^{\mathrm{U}}) = (45, 0.7)\) | ||
\(V_{50} < 15\%\) | \((D_{2}^{\mathrm{U}},\zeta_{2}^{\mathrm{U}}) = (50, 0.85)\) | ||
OAR3 (cyan) | Rectum | \(V_{20} < 40\%\) | \((D_{2}^{\mathrm{U}},\zeta_{2}^{\mathrm{U}}) = (20, 0.6)\) |
\(V_{30} < 20 \%\) | \((D_{2}^{\mathrm{U}},\zeta_{2}^{\mathrm{U}}) = (30, 0.8)\) | ||
\(V_{35} < 10\%\) | \((D_{2}^{\mathrm{U}},\zeta_{2}^{\mathrm{U}}) = (35, 0.9)\) | ||
OAR4 (magenta) | Left femoral head | \(V_{30} < 10\%\) | \((D_{2}^{\mathrm{U}},\zeta_{2}^{\mathrm{U}}) = (30, 0.9)\) |
For the phantom imitating prostate cancer, in accordance with the Memorial Sloan Kettering Cancer Center (MSKCC) protocol of the report [30, 31], we applied five-field irradiation at angles of 45, 135, 195, 270, and 345 degrees; these angles were measured counterclockwise from the horizontal line that passes through the center of the CT image in Figure 1. Since we should find that the intensity of radiation beams affects only the PTV and OARs, we obtain \(J = 806\). In the calculation of K, we did not take scattering of radiation [34] into consideration to simplify the dose calculation.
The IMRT planning derived from the above mentioned construction is inconsistent and acceptable in the meaning of the definitions in Section 2.
5 Experimental results
As with conventional IMRT planning optimization, the method of the split feasibility problem as well as the gradient methods such as Newton’s method and the conjugate gradient method are known to be used for objective function minimization. All of these methods for searching for feasibility solutions are based on iterative procedure. However, the differential equation described in Eq. (11) is a continuous analog of the CQ-algorithm [26], which is designed for implementing the split feasibility problem with the advantage of not calculating a matrix inversion or Hessian over the conventional gradient methods, and is suitable for comparing and verifying the mechanism of obtaining acceptable solutions to the proposed continuous system in Eq. (8).
The common parameters and conditions for numerical simulation are as follows. For numerical integration of differential equations, we used the solver ode15s provided by MATLAB (MathWorks, Natick, USA). The parameter γ was set to 100, and the initial values of the state variables at \(t = 0\) were chosen as \(x_{j}^{0} = 10\) for any j. The upper beam limit γ is defined to avoid machine restrictions, and we are able to confirm that the IMRT planning system cannot reach any acceptable solution if the second inequality of Eq. (4) with x replaced by the vector containing each element of γ is not satisfied.
6 Discussion and conclusion
The purpose of this paper is to present a dose-volume constrained optimization method of IMRT treatment planning using the initial-value problem of the continuous-time dynamical system. First, let us discuss the theoretical results obtained by using the proposed method. Theorem 1 guarantees that the value of the Lyapunov function W, which corresponds to an objective or cost function, monotonically decreases along a solution obtained by the proposed method when an IMRT planning system is acceptable, which is mathematically defined for the existence of radiation beams satisfying the given DVCs. Additionally, when we choose an initial state in Ω, our method is able to find a box-constrained solution within Ω that is supported by Proposition 1.
Then we discuss the convergence to an equilibrium in the acceptable set of solutions in numerical experiments. The behavior can be explained qualitatively as follows. When the dose distribution for a PTV or an OAR becomes partly acceptable as time passes, the value of the corresponding function Q and the term added to the vector field (the right-hand side of Eq. (8)), which is required to make it acceptable, are zero in the dynamical system. Therefore, it is expected that the dynamics will make it easier for other OARs and PTVs to work to be partly acceptable. The function Q plays an essential role to control the cooperation among terms for DVCs in the vector field. If all dose distributions in PTVs and OARs become partly acceptable, then the vector field is entirely zero after the solution has converged to an equilibrium corresponding to the desired radiation beam weights.
In accordance with the theoretical result, we can take any initial state \(x^{0}\) in Ω. Indeed, when performing numerical experiments with several other initial values \(x_{j}^{0}\) for all j in the range \((0, \gamma)\), similar results of convergence to equilibria in the acceptable set \({\mathcal{A}}\) were obtained for all trials. Note that our goal is to reach an arbitrary element in \({\mathcal{A}}\), although an optimal solution is, in general, an isolated point in the state space in the usual definition of optimization problems. Nevertheless, it is significant to consider an optimization problem restricted to the set \({\mathcal{A}}\) from the viewpoint of another object. Regarding this, the following property of solutions was observed experimentally: the mean value of the state variables in an acceptable equilibrium after obtaining convergence was relatively small when the initial state \(x_{j}^{0}\) was taken as a small value. In practical IMRT planning, the radiation beam weights are expected to be as weak as possible for reducing normal tissue doses. Hence, we can say that we should choose a small initial value so long as the selection of the initial state in Ω results in the same convergence to an element in \({\mathcal{A}}\).
A discretization of the differential equation in Eq. (11) using the Euler method leads to the iterated CQ-algorithm; therefore, the continuous system has the same convergence property on solutions as that of the difference equation. That is, the continuous system possesses a good characteristic that its solutions converge asymptotically to an equilibrium if the given inverse problem is consistent; otherwise, a close solution to the complete solution can be found [35]. By extending the continuous system, we have presented a hybrid dynamical system that can change the vector fields depending on partly acceptable conditions. For the inverse problem dealing with the discontinuous constraints defined by proportion rates, the global stability of solutions of the hybrid system can be proved theoretically using the tool of the Lyapunov theorem.
Although the proposed method of using dynamical systems with piecewise continuous vector fields is well designed for dose-volume constrained optimization, it requires numerical integration with a high computational cost to solve solutions to differential equations. In future work, we will attempt to construct an iterative method formulated by discretizing our differential equations for reducing the costs.
Declarations
Acknowledgements
This research was partially supported by JSPS KAKENHI Grant Number 24560522.
Open Access This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly credited.
Authors’ Affiliations
References
- Palta, JR, Mackie, TR (eds.): Intensity-Modulated Radiation Therapy: The State of the Art. Med. Phys., Madison (2003) Google Scholar
- Bortfeld, T: IMRT: a review and preview. Phys. Med. Biol. 51(13), 363-379 (2006) View ArticleGoogle Scholar
- Shao, L: A survey of beam intensity optimization in IMRT. In: Halliburton, T (ed.) Proceeding of the 40th Annual Conference of the Operational Research Society of New Zealand, pp. 255-264. The Operational Research Society of New Zealand, Wellington (2005) Google Scholar
- Wu, Q, Mohan, R: Algorithms and functionality of an intensity modulated radiotherapy optimization system. Med. Phys. 27(4), 701-711 (2000) View ArticleGoogle Scholar
- Cotrutz, C, Lahanas, M, Kappas, C, Baltas, D: A multiobjective gradient-based dose optimization algorithm for external beam conformal radiotherapy. Phys. Med. Biol. 46(8), 2161-2175 (2001) View ArticleGoogle Scholar
- Wu, Q, Mohan, R: Multiple local minima in IMRT optimization based on dose-volume criteria. Med. Phys. 29(7), 1514-1527 (2002) View ArticleGoogle Scholar
- Lahanas, M, Schreibmann, E, Baltas, D: Multiobjective inverse planning for intensity modulated radiotherapy with constraint-free gradient-based optimization algorithms. Phys. Med. Biol. 48(17), 2843-2871 (2003) View ArticleGoogle Scholar
- Zhang, X, Liu, H, Wang, X, Dong, L, Wu, Q, Mohan, R: Speed and convergence properties of gradient algorithms for optimization of IMRT. Med. Phys. 31(5), 1141-1152 (2004) View ArticleGoogle Scholar
- Bortfeld, T, Kufer, KH, Monz, M, Trofimov, A, Niemierko, A: Problems with current IMRT prescription practices and planning systems (abstract). Med. Phys. 31, 1761 (2004) Google Scholar
- Gavurin, MK: Nonlinear functional equations and continuous analogies of iterative methods. Izv. Vysš. Učebn. Zaved., Mat. 5(6), 18-31 (1958) MathSciNetGoogle Scholar
- Schropp, J: Using dynamical systems methods to solve minimization problems. Appl. Numer. Math. 18(1), 321-335 (1995) View ArticleMATHMathSciNetGoogle Scholar
- Airapetyan, RG: Continuous Newton method and its modification. Appl. Anal. 73(3-4), 463-484 (1999) View ArticleMATHMathSciNetGoogle Scholar
- Airapetyan, RG, Ramm, AG, Smirnova, AB: Continuous analog of Gauss-Newton method. Math. Models Methods Appl. Sci. 9(3), 463-474 (1999) View ArticleMATHMathSciNetGoogle Scholar
- Airapetyan, RG, Ramm, AG: Dynamical systems and discrete methods for solving nonlinear ill-posed problems. In: Anastassiou, GA (ed.) Applied Mathematics Reviews, vol. 1, pp. 491-536. World Scientific, Singapore (2000) View ArticleGoogle Scholar
- Ramm, AG: Dynamical systems method for solving operator equations. Commun. Nonlinear Sci. Numer. Simul. 9(4), 383-402 (2004) View ArticleMATHMathSciNetGoogle Scholar
- Li, L, Han, B: A dynamical system method for solving nonlinear ill-posed problems. Appl. Math. Comput. 197(1), 399-406 (2008) View ArticleMATHMathSciNetGoogle Scholar
- Fujimoto, K, Al-Ola, OA, Yoshinaga, T: Continuous-time image reconstruction using differential equations for computed tomography. Commun. Nonlinear Sci. Numer. Simul. 15(6), 1648-1654 (2010) View ArticleMATHGoogle Scholar
- Al-Ola, OMA, Fujimoto, K, Yoshinaga, T: Common Lyapunov function based on Kullback-Leibler divergence for a switched nonlinear system. Math. Probl. Eng. 2011, Article ID 723509 (2011) Google Scholar
- Yamaguchi, Y, Fujimoto, K, Al-Ola, OMA, Yoshinaga, T: Continuous-time image reconstruction for binary tomography. Commun. Nonlinear Sci. Numer. Simul. 18(8), 2081-2087 (2013) View ArticleMATHMathSciNetGoogle Scholar
- Fujimoto, K, Tanaka, Y, Al-Ola, OMA, Yoshinaga, T: Continuous-time method and its discretization to inverse problem of intensity-modulated radiation therapy treatment planning. Commun. Nonlinear Sci. Numer. Simul. 19(6), 1996-2004 (2014) View ArticleMathSciNetGoogle Scholar
- Leine, RI, van de Wouw, N: Stability and Convergence of Mechanical Systems with Unilateral Constraints. Lecture Notes in Applied and Computational Mechanics, vol. 36. Springer, Berlin (2008) MATHGoogle Scholar
- Lyapunov, AM: Stability of Motion. Academic Press, New York (1966) MATHGoogle Scholar
- Kullback, S, Leibler, RA: On information and sufficiency. Ann. Math. Stat. 22(1), 79-86 (1951) View ArticleMATHMathSciNetGoogle Scholar
- Csiszár, I: Why least squares and maximum entropy? An axiomatic approach to inference for linear inverse problems. Ann. Stat. 19(4), 2032-2066 (1991) View ArticleMATHGoogle Scholar
- Censor, Y, Elfving, T: A multiprojection algorithm using Bregman projections in a product space. Numer. Algorithms 8(2), 221-239 (1994) View ArticleMATHMathSciNetGoogle Scholar
- Byrne, CL: Iterative oblique projection onto convex sets and the split feasibility problem. Inverse Probl. 18, 441-453 (2002) View ArticleMATHMathSciNetGoogle Scholar
- Deepho, J, Kumam, P: Split feasibility and fixed-point problems for asymptotically quasi-nonexpansive mapping. J. Inequal. Appl. 2013, 322 (2013) View ArticleMathSciNetGoogle Scholar
- Hou, Q, Wang, J, Chen, Y, Galvin, JM: An optimization algorithm for intensity modulated radiotherapy - the simulated dynamics with dose-volume constraints. Med. Phys. 30(1), 61-68 (2003) View ArticleGoogle Scholar
- The Visible Human Project. http://www.nlm.nih.gov/research/visible/visible_human.html. Accessed 24 Oct 2014
- Lu, R, Radke, RJ, Yang, J, Happersett, L, York, E, Jackson, A: Reduced-order constrained optimization in IMRT planning. Phys. Med. Biol. 53(23), 6749-6766 (2008) View ArticleGoogle Scholar
- Cahlon, O, Zelefsky, MJ, Shippy, A, Chan, H, Fuks, Z, Yamada, Y, Hunt, M, Greenstein, S, Amols, H: Ultra-high dose (86.4 Gy) IMRT for localized prostate cancer: toxicity and biochemical outcomes. Int. J. Radiat. Oncol. Biol. Phys. 71(2), 330-337 (2008) View ArticleGoogle Scholar
- Pederson, AW, Fricano, J, Correa, D, Pelizzari, CA, Liauw, SL: Late toxicity after intensity-modulated radiation therapy for localized prostate cancer: an exploration of dose-volume histogram parameters to limit genitourinary and gastrointestinal toxicity. Int. J. Radiat. Oncol. Biol. Phys. 82(1), 235-241 (2012) View ArticleGoogle Scholar
- Quan, EM, Li, X, Li, Y, Wang, X, Kudchadker, RJ, Johnson, JL, Kuban, DA, Lee, AK, Zhang, X: A comprehensive comparison of IMRT and VMAT plan quality for prostate cancer treatment. Int. J. Radiat. Oncol. Biol. Phys. 83(4), 1169-1178 (2012) View ArticleGoogle Scholar
- Kawrakow, I: Improved modeling of multiple scattering in the voxel Monte Carlo model. Phys. Med. 4, 505-517 (1997) View ArticleGoogle Scholar
- Censor, Y, Bortfeld, T, Martin, B, Trofimov, A: A unified approach for inversion problems in intensity-modulated radiation therapy. Phys. Med. Biol. 51, 2353-2365 (2006) View ArticleGoogle Scholar