Author: Yang Sheng, Qingrong Jackie Wu, Qiuwen Wu, Xin Wu, Dongrong Yang 👨🔬
Affiliation: Duke University Medical Center 🌍
Purpose: Gradient-based optimization is the standard approach for IMRT fluence map optimization (FMO). Recently, a machine learning (ML)-assisted framework using a one-layer neural network was proposed and shown to be feasible for FMO. However, its characteristics were not fully understood. This study aims to provide a detailed quality and performance evaluation of the ML-assisted framework in comparison to the standard gradient-based approach.
Methods: The classical steepest gradient descent optimizer was selected as the benchmark for reference and comparison. Both optimizers utilized identical inputs and objective functions to ensure a fair evaluation. DVH and gEUD based objectives were implemented in standard quadratic forms. The ML-assisted framework was developed using Pytorch’s L-BFGS optimizer with GPU acceleration. Quality metrics included final objective values, DVHs, fluence maps, and dose distributions, while performance was primarily assessed based on the number of iterations to convergence. The optimizations were conducted on prostate and head-and-neck cases.
Results: The ML-assisted framework demonstrated comparable or superior quality compared to classical optimization, notably outperforming it in many cases by achieving lower objective values and improved DVHs. In terms of efficiency, the ML-assisted framework required significantly less iterations and time to achieve similar or better results. Across 43 prostate case runs, classical optimization required 77.39±11.77 iterations to reduce the initial objective by 99%, whereas ML-assisted framework achieved the same reduction in just 1.30±0.83 iterations. Similarly, for 57 runs using gEUD-based constraints, the iterations were 154.99±9.89 for classical and 0.88±0.04 for all ML-assisted optimizations, indicating that a single ML-assisted iteration could be equivalent to approximately 100 classical iterations.
Conclusion: The ML-assisted framework presents a robust and computationally efficient alternative to classical optimization methods for FMO, delivering faster convergence and enhanced navigation of complex optimization spaces while maintaining or exceeding solution quality.