High-dimensional automated radiation therapy treatment planning via Bayesian optimization.

Area of Science:

• Medical Physics
• Radiation Oncology
• Computational Optimization

Background:

• Radiation therapy treatment planning involves balancing competing clinical objectives.
• Automated planning can streamline this complex hyperparameter tuning process.
• High-dimensional settings pose challenges for traditional optimization methods.

Purpose of the Study:

• To evaluate modern Bayesian optimization (BO) methods for automated treatment planning.
• To compare the performance of standard (GPEI) and high-dimensional (SAAS-BO) BO techniques.
• To assess automated plans against clinical plans and other optimization baselines.

Main Methods:

• Retrospective analysis of 20 locally advanced rectal cancer patients treated with IMRT.
• Implementation of an automated planning framework using Gaussian Process with Expected Improvement (GPEI) and Sparse Axis Aligned Subspace BO (SAAS-BO).
• Comparison with Nelder-Mead simplex search and random tuning, evaluating plan quality and efficiency against clinical plans.

Main Results:

• SAAS-BO plans demonstrated comparable target hot spot control and homogeneity to clinical plans, outperforming GPEI and Nelder-Mead.
• Both SAAS-BO and GPEI plans significantly improved conformity and reduced dose spillage compared to clinical plans.
• All automated methods reduced dosimetric indices for the femoral head and bladder; BO methods identified sensitive planning parameters.

Conclusions:

• A BO-based framework successfully automates treatment planning, yielding high-quality results.
• Tested BO methods (GPEI, SAAS-BO) produce superior plans and decrease planner workload.
• Analysis confirms the inherent low dimensionality of these treatment planning problems.