Premium
TH‐AB‐BRB‐01: Solving Larger IMRT Problems by Enhanced Reduced‐Order Constrained Optimization
Author(s) -
Nourzadeh H,
Radke R,
Jackson A,
Bakr S,
Tuomaala S
Publication year - 2015
Publication title -
medical physics
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 1.473
H-Index - 180
eISSN - 2473-4209
pISSN - 0094-2405
DOI - 10.1118/1.4926132
Subject(s) - mathematical optimization , optimization problem , algorithm , sampling (signal processing) , pipeline (software) , mathematics , principal component analysis , computer science , artificial intelligence , filter (signal processing) , computer vision , programming language
Purpose: To report the latest improvements in the pipeline of the reduced‐order constrained optimization (ROCO) method for fast IMRT planning. Methods: The ROCO method involves the characterization of the solution space of the underlying fluence map IMRT optimization problem in a compact form, which enables forming a clinically tractable constrained optimization problem. The offline part of the method consists of three stages: sampling, learning the effective mode space, and mode‐dose calculation. In the sampling stage, a set of unconstrained optimization problems is solved, each corresponding to a different choice of weights and dose‐volume limits. As opposed to previous implementations, we preprocess the samples, identify the effective beamlets with non‐zero variances, and use them to attain a compact representation of the fluence space using Principal Component Analysis (PCA). In the mode‐dose calculation stage, the dose corresponding to each mode is computed based on the underlying full dose‐calculation technique. We offer a new formulation for handling possible negative mode elements, exploiting linearity of the mode‐dose map. In the online phase, a constrained optimization problem with a minimal number of constraints is solved over the basis coefficients spanning the reduced‐size space. Results: The results suggest that for a typical prostate case, the PCA stage can be performed about 5 times faster. Moreover, the mode dose calculation performance is accelerated by a factor of 2. Apart from computational efficiency, we also observed significantly lower memory demand for ROCO model compared to optimization models relying on pre‐calculation of dose deposition coefficients (DDCs). The proposed method has been integrated into the Varian Eclipse version 13.5 treatment planning system as a stand‐alone application. Conclusion: The new implementation has led a substantial savings in memory requirements and computational effort, allowing larger IMRT problems with finer beamlet resolution to be addressed with the available memory resources in less time.