Author: Ara Alexandrian, Sadiki Daniel 👨🔬
Affiliation: Louisiana State University, Mary Bird Perkins Cancer Center 🌍
Purpose: To develop a learning-to-optimize machine learning model that accelerates optimization in VMAT treatment planning by training on prostate patient data.
Methods: A treatment plan dataset of fifty VMAT prostate patients were input into MatRad—an open-source multi-modality radiation treatment planning system—for dose calculation and optimization using interior point optimization(IPOPT). At each iteration of the IPOPT, the corresponding objective function values were recorded along with their associated features extracted from the linear accelerator (LINAC) control points. A Learning to Optimize (L2O) framework was applied, utilizing a neural network to train the collected data. The model was trained to identify optimization patterns by learning the relationship between the objective function values and their corresponding features for all patients.
Results: This research is ongoing, and the implementation of the Learning to Optimize (L2O) framework is expected to enhance the efficiency of VMAT plan optimization. By leveraging a neural network trained on objective function values and their associated LINAC control point features, the model aims to predict optimal solutions and reduce the number of required iterations for optimization. Preliminary expectations suggest that this approach could accelerate the optimization process, potentially reducing computational time by 40-50% while maintaining or improving dose quality.
Conclusion: These anticipated improvements would streamline treatment planning, enabling faster plan adaptation and increased clinical workflow efficiency. By predicting optimal solutions from objective function values and LINAC control point features, this approach aims to reduce computational time while maintaining dose quality. Future work will focus on validating the model’s performance against conventional methods that are currently implemented to optimize plans, in order to assess its quantifiable impact on computational speed and treatment quality. Additionally, evaluations of the improvement in dose quality as a result of the applied L2O model would be made to critique the feasibility of its performance.