Abstract
Radiation therapy treatment planning can be viewed as an iterative hyperparameter tuning process to balance conflicting clinical goals. In this work, we investigated the performance of modern Bayesian Optimization (BO) methods on automated treatment planning problems in high-dimensional settings. 20 locally advanced rectal cancer patients treated with intensity-modulated radiation therapy (IMRT) were retrospectively selected as test cases. We implemented an automated treatment planning framework that adjusts dose objectives and weights simultaneously, and tested the performance of two BO methods on the treatment planning task: one standard BO method (GPEI) and one BO method dedicated to high-dimensional problems (SAAS-BO). A random tuning method was also included as the baseline. We compared and analyzed the three automated methods' plan quality and planning efficiency and the different search patterns of the two BO methods. For the target structures, the SAAS-BO plans achieved comparable hot spot control ($p=0.43$) and homogeneity ($p=0.96$) with the clinical plans, significantly better than the GPEI and random plans ($p<0.05$). Both SAAS-BO and GPEI plans significantly outperformed the clinical plans in conformity and dose spillage ($p<0.05$). Compared with clinical plans, the treatment plans generated by the three automated methods all made reductions in evaluated dosimetric indices for the femoral head and the bladder. The analysis of the underlying predictive models has shown that both BO procedures have identified similar important planning parameters. This work implemented a BO-based hyperparameter tuning framework for automated treatment planning. Both tested BO methods were able to produce high-quality treatment plans and the model analysis also confirmed the intrinsic low dimensionality of the tested treatment planning problems.
Introduction
In radiation therapy, "planning" is a high-stakes balancing act. Clinicians must deliver enough radiation to destroy a tumor while meticulously sparing the surrounding healthy organs. Traditionally, this involves a human planner manually "turning knobs"—adjusting dozens of dose objectives and weights in a trial-and-error process. It is time-consuming, subjective, and increasingly difficult as the number of adjustable parameters grows.
Inverse planning requires setting "hyperparameters"—the weights and goals for an optimization engine. For a typical rectal cancer case, a planner might need to manage over 30 different parameters simultaneously. Standard optimization techniques often struggle with this "curse of dimensionality." Furthermore, evaluating a single "guess" at these parameters requires running a full dose simulation, which is computationally expensive. This research addresses how to efficiently find the optimal "knob settings" in a high-dimensional space without needing thousands of expensive simulations or a library of previous patient data.
Methods
The researchers developed an automated framework that uses Bayesian Optimization (BO) to navigate the complex, high-dimensional parameter space of radiation therapy planning. The framework was integrated with a commercial Treatment Planning System (Eclipse TPS).
The study compared three different search strategies:
- GPEI (Standard BO): Uses a Gaussian Process (GP) as a surrogate model and Expected Improvement (EI) to pick the next set of parameters.
- SAAS-BO (Sparse Axis Aligned Subspace BO): A modern BO variant designed for high dimensions. It uses a "sparsity-inducing prior," assuming that out of many parameters, only a few are truly critical to the plan's success.
- Random Search: A baseline method using a Sobol sequence to sample the parameter space.
The "score" for each plan was determined by a custom Score Function—a weighted sum of multiple Plan Quality Metrics (PQMs) like target coverage, organ-at-risk (OAR) sparing, and dose fall-off.
Key Results and Insights
The study used a retrospective dataset of 20 locally advanced rectal cancer patients with a 34-dimensional input space. A significant finding was confirming that radiation therapy planning has intrinsic low dimensionality. Even with 34 parameters, the SAAS-BO model correctly identified that a small subset (primarily related to the bladder and "avoidance" structures) drove most of the plan quality.
Both BO methods generated plans that significantly outperformed clinical plans in conformity (how tightly the dose hugs the tumor) and dose spillage (preventing radiation from "leaking" into healthy tissue). SAAS-BO achieved comparable hot spot control ($p=0.43$) and homogeneity ($p=0.96$) to expert human plans, whereas standard GPEI and Random search were significantly worse ($p<0.05$). All automated methods reduced the dose to the femoral heads and bladder compared to clinical plans.
In terms of efficiency, SAAS-BO converged to a high-quality plan faster than GPEI. However, the computational overhead of SAAS-BO (using MCMC sampling) meant it took longer per iteration (~7 hours for a full 120-iteration run vs ~3 hours for GPEI).
Discussion
While the planning is automated, the total time remains a hurdle for clinical environments. Future work may involve parallel BO or early stopping to mitigate this. Additionally, the current model treats the score function as a "black box." The authors suggest that "Composite Function BO"—which looks inside the score function to see how individual organs are faring—could lead to even more efficient optimization. Future studies will look to expand this to more complex sites like Head & Neck or Lung cancers.
Summary updated