Skip to main content
eScholarship
Open Access Publications from the University of California

Fully Automated Radiation Therapy Treatment Planning Through Knowledge-Based Dose Predictions

  • Author(s): Landers, Angelia
  • Advisor(s): Sheng, Ke
  • et al.
Abstract

Intensity-modulated radiotherapy treatment planning is an inverse problem that typically includes numerous parameters that have to be manually tuned by expert planners. This process can take hours or even days and can often lead to suboptimal plans. In this study, we developed a technique for fully automated radiotherapy treatment planning with the guidance of dose predictions using high quality or evolving knowledge bases.

Knowledge-based planning (KBP) dose prediction provides patient-specific estimations for the capabilities and limitations of a plan. Statistical voxel dose learning (SVDL) was developed to predict the voxel dose of new patients. The method was compared to supervised machine learning methods, spectral regression (SR) and support vector regression (SVR), to evaluate the prediction accuracy and robustness of using small training sets. SVDL was found to have higher prediction accuracy than the more sophisticated machine learning methods and effective even with small training sets.

To remove any dependence on hyperparameters that require manual tuning, voxel-based non-coplanar 4π radiotherapy and coplanar volumetric modulated arc therapy (VMAT) optimization problems were modified to include the KBP predicted doses. The new cost functions encourage the plans to meet or improve on the predicted doses. Because of this, the resulting plan quality is heavily reliant on the plan quality of the KBP training set. To ensure high quality plans, non-coplanar and coplanar IMRT plans were manually created using all available beams. The resulting automated plans were of superior quality compared to manually-created plans.

In the case of no existing high quality training set, evolving-knowledge-base (EKB) planning was developed. An initial, low quality training set was used for the first epoch of automated planning. In subsequent epochs, the superior plans from the previous epoch were taken as the training set. Overall plan quality was observed to improve through epochs, plateauing after 3 and 6 epochs for lung and head & neck planning, respectively. The final EKB plans were significantly higher quality than manually-created VMAT plans and equivalent to manually-created 4π plans.

Through the course of this work, we established a robust and accurate KBP dose prediction technique, which we then utilized in our automated planning protocol. Both the use of high quality training sets and EKB planning created high quality plans in a more efficient and consistent manner than hyperparameter tuning.

Main Content
Current View