Fluence optimization
The goal of the fluence optimization is to find a set of bixel/spot weights that yield the best possible dose distribution according to the clinical objectives and constraints underlying the radiation treatment.
For mathematical optimization, these clinical objectives and constraints have to be translated into mathematical objectives and constraints. matRad supports the mathematical optimization of a weighted sum of objectives to help finding an optimal trade-off between adequate target coverage and normal tissue sparing for an individual patient as well as the formulation of constraints. The individual objectives and constraints are defined per structure, can be chosen by the user and include, among others:
- squared overdosage objective
- squared underdosage objective
- squared dose deviation objective
- mean dose objective
- equivalent uniform dose objective
- min/max dose constraints
- DVH point objectives and constraints
The overall fluence optimization process is coordinated by the matRad function matRad_fluenceOptimization.m.
Since version 2.10.0 objectives and constraints are implemented with an object oriented approach. For the old format, see further down below.
At the moment, matRad allows for objectives and constraints based on dose. Each objective and constraint is defined in a class, found in the optimization/+DoseObjectives & optimization/+DoseConstraints. This enables the easy implementation of new (dose-based) constraints & objectives by writing your own class, inheriting from matRad_DoseObjective or matRad_DoseConstraint, and implementing the therein declared interface, i.e. the parameter definition, and respective objective/constraint function and its gradient/jacobian.
New objectives/constraints are then automatically recognized also in the GUI.
This is extended to the implementation of optimization problems and optimizers. An optimization problem combines the single objectives into an objective function and organizes the constraint structure to give a standard optimization problem to the optimizer. matRad defaults to the IPOPT package for large scale non-linear optimization which is included via a MEX file and interfaced in matRad_OptimizerIPOPT. MATLAB's proprietary fmincon support is added in matRad_OptimizerFmincon, other optimizers may be added as well. So far, only non-linear constrained optimization is supported by matRad_OptimizationProblem and for optimizers.
Optimizers can be changed by setting pln.propOpt.optimizer.
The matRad_OptimizationProblem class also enables to implement advanced planning problems, like direct aperture optimization as implemented in matRad_OptimizationProblemDAO.
The objectives and constraints are stored as an dose objective struct within the cst cell. The objectives and constraints can be set including all necessary parameters via the matRad GUI. All functions involved in the optimization process are located in a subfolder called "optimization" within the matRad root folder. matRad relies on the IPOPT package for large scale non-linear optimization which is included via a MEX file. IPOPT requires call back functions for objective function, gradient, constraint, and Jacobian evaluation. We use the wrapper functions matRad_objFuncWrapper.m, matRad_gradFuncWrapper.m, matRad_constFuncWrapper.m, and matRad_jacobFuncWrapper.m to coordinate the evaluation of all defined objectives and constraints.
All optimization functionalities work equally for optimization processes based on physical dose as well as biological effect Wilkens & Oelfke (2006) and RBE-weighted dose according to Krämer & Scholz (2006). The biological effect and the RBE-weighted dose are calculated with α and β base data that has been calculated according to the local effect model IV. α and β tables are available as part of the base data set carbon_Generic which is provided with the matRad release.
For photons, matRad also features an experimental direct aperture optimization that largely follows the implementation described in Wild et al. (2015) which is based on Bzdusek et al. (2009) and (with some modification) Unkelbach & Cassioli (2012).
