Click here to


Are you sure ?

Yes, do it No, cancel

A Computationally Efficient Algorithm for Integrating Dose-Volume Constraints Into IMRT Fluence Optimization for Automated Treatment Planning

S Mukherjee*, L Hong , J Deasy , M Zarepisheh , Memorial Sloan Kettering Cancer Center, New York, NY


(Monday, 7/30/2018) 4:30 PM - 5:30 PM

Room: Exhibit Hall | Forum 3

Purpose: To develop a computationally efficient technique to handle dose-volume constraints (DVCs) for IMRT fluence map optimization.The technique is adopted by our in-house automated treatment planning system.

Methods: Dose-volume constraints, in principle, can be incorporated into optimization by introducing some “binary� variables. For each voxel, the corresponding binary variable is either “0� or “1� indicating whether the voxel meets or violates the DVC threshold. A constraint to limit the number of voxels violating the DVC threshold can be imposed to enforce the DVC. However, this optimization problem, known as mixed integer programming (MIP) problem, is non-convex and computationally expensive. We propose solving a convex relaxation problem by replacing the binary variables with the regular continuous variables that simply requires the variables to be between 0 and 1. Although this convex relaxation does not guarantee the DVC satisfaction, it provides crucial initial information about organ-at-risk (OAR) voxels that receive low-dose radiation. We can then impose maximum dose constraints on that low-dose region to enforce DVCs. The proposed technique is adopted by our in-house automated treatment planning system which is based on hierarchical constrained optimization. The PTV coverage is maximized in the objective function while the DVCs on OAR are enforced as hard constraints.

Results: The proposed algorithm is successfully tested on a series of patients with paraspinal tumors. For one patient for example, the computational time was 5.2 and 5.1 minutes with and without DVC respectively, indicating the new algorithm only added 2% to the computational time to satisfy DVCs. We were unable to solve the non-convex MIP problem after several days.

Conclusion: We developed and implemented a new algorithm for fluence map optimization that incorporates DVCs into the optimization as hard constraints. The algorithm was computationally fast and capable of satisfying DVCs without any parameter tweaking.


Not Applicable / None Entered.


TH- External beam- photons: IMRT dose optimization algorithms

Contact Email