Latent space arc therapy optimization

Area of Science:

• Medical Physics
• Radiation Oncology
• Computational Optimization

Background:

• Volumetric modulated arc therapy (VMAT) planning involves complex, high-dimensional, non-convex optimization.
• Current methods use heuristics for initialization, leading to slower optimization and favoring local optima due to plan overparameterization.

Purpose of the Study:

• To address VMAT plan overparameterization and accelerate treatment planning.
• To reduce the effective dimensionality of VMAT treatment plans.

Main Methods:

• Utilized unsupervised deep learning to reduce the dimensionality of VMAT treatment plans.
• Developed an optimization engine based on these low-dimensional arc representations.

Main Results:

• Successfully reduced the effective dimension of VMAT treatment plans.
• Facilitated faster radiation therapy planning times through optimized low-dimensional representations.

Conclusions:

• Unsupervised deep learning offers a viable method to overcome VMAT planning challenges.
• This approach enhances optimization efficiency, leading to quicker and potentially improved radiation treatment plans.