A machine learning toolkit assisted approach for IMRT fluence map optimization: feasibility and advantages.

Area of Science:

• Medical Physics
• Computational Biology
• Radiotherapy

Background:

• Traditional machine learning (ML) and deep learning (DL) in treatment planning often require complex models and extensive datasets, limiting their ability to fully replace conventional optimization.
• Existing ML/DL applications have not entirely supplanted established optimization processes in radiotherapy treatment planning.

Purpose of the Study:

• To present a novel application of ML, utilizing established ML/DL toolkits directly for treatment plan optimization.
• To evaluate the efficacy of ML-assisted treatment plan optimization against classical methods.

Main Methods:

• A one-layer neural network was designed using the dose deposition matrix and implemented with PyTorch's L-BFGS optimizer, leveraging GPU acceleration.
• Compared the MLT-assisted framework against the classical steepest descent optimizer using identical inputs and objective functions (DVH- and gEUD-based).
• Tested optimizer performance across various initial conditions, including uniform and 1,000 random starting points for prostate and head-and-neck cancer cases.

Main Results:

• The MLT-assisted framework achieved comparable or superior plan quality, evidenced by lower objective values, improved dose-volume histograms (DVHs), and finer fluence map details.
• Demonstrated faster convergence, requiring significantly fewer iterations and evaluations than classical optimization, especially for gEUD-based objectives.
• Showed increased robustness against initial conditions, being less prone to getting trapped in suboptimal solutions compared to classical methods.

Conclusions:

• The developed MLT-assisted framework offers a novel approach to treatment plan optimization, leveraging ML toolkits for enhanced efficiency and reliability.
• This method enables faster convergence, greater robustness, and effective handling of complex constraints in radiotherapy planning.
• Establishes MLT-assisted optimization as a practical and effective alternative to conventional optimization techniques in radiation therapy.